CIKM 2026 Conference Program

Learning generalizable road representations is essential for urban computing, yet remains challenging due to heterogeneous urban signals, city-specific road characteristics, and diverse downstream tasks. Existing methods often encode multi-source information jointly, causing functional, structural, and mobility semantics to be entangled in a single representation space. We propose UniRoad, a unified multi-view framework for cross-city road representation learning. UniRoad decomposes road semantics into three complementary views: an LLM-based profile view that captures functional semantics from road attributes and POI contexts, a prototype-based structure view that learns transferable structural roles from local topology, and a trajectory-informed mobility view that models travel patterns from road sequences. UniRoad further introduces confidence-guided fusion to adaptively integrate these views according to their reliability. Extensive experiments on four cities and four downstream tasks demonstrate that UniRoad consistently outperforms strong baselines and achieves robust cross-city transferability. Code and data are available at: https://github.com/csjiezhao/UniRoad.

Geospatial representation learning from satellite imagery is a fundamental problem for large-scale urban analysis and real-world applications. Despite recent advances, current methods struggle with cross-region generalization and semantic interpretability due to their reliance on region-specific auxiliary data and the neglect of semantic alignment within multi-temporal urban imagery. Therefore, we present CoST, a novel \underline{Co}ntrastive-based \underline{S}patial-\underline{T}emporal framework that aligns spatial context with multi-temporal semantics to extract universal geographic regularities shared across regions. Specifically, CoST explicitly models spatial correlations to capture transferable geographic structures and exploits multi-year urban change semantics to align learned representations with high-level geo-semantics. Extensive experiments demonstrate that CoST consistently achieves superior performance across various downstream tasks and in unseen scenario, yielding an average relative gain of 8.7\% over the strongest competing methods across eight city-indicator settings. The code is available in \href{https://github.com/Arandinglv/CoST}{this repo}.

Urban spatiotemporal prediction is crucial for smart city management, yet fragmented data across administrative boundaries significantly impedes effective model training. To overcome data limitations and leverage knowledge from data-rich regions while preserving data privacy, federated transfer learning has emerged as a promising approach for cross-regional urban prediction tasks. However, traditional learning methods encounter the dual challenge: (1) intra-regional knowledge acquisition difficulties due to missing boundaries that weaken spatiotemporal dependencies and cause data sparsity; and (2) cross-regional knowledge transfer bias, where distribution differences between source and target regions trigger negative transfer and reduce model robustness. To address these issues, we propose CRAFT: a novel federated transfer learning framework that integrates Large Language Models (LLMs) for urban spatiotemporal prediction. Leveraging LLMs' powerful semantic understanding and contextual reasoning abilities, CRAFT captures complex spatiotemporal dependencies and uncovers latent regional connections from heterogeneous data sources, thus alleviating boundary-induced data sparsity. Additionally, to address inter-regional data heterogeneity, we employ a mixture-of-experts (MoE) model with parameter decoupling, which excels at handling diverse data distributions while minimizing federated communication costs. Experimental results on real-world datasets from multiple cities demonstrate the effectiveness of CRAFT in cross-regional federated transfer learning. The code is publicly available at: https://github.com/Jasonxx4/CRAFT.

Spatiotemporal forecasting requires modeling temporal dynamics and spatial dependencies across multiple spatial units. Existing approaches construct spatial relations from graph structures, latent representations, or attention weights. This work uses frequency-bin-level temporal responses in the prototype routing cost before spatial-unit-level aggregation. We propose BPR-Net (Bin-aware Prototype Routing Network), a frequency-bin-aware prototype routing model for spatiotemporal forecasting. BPR-Net explicitly represents frequency-bin-level temporal responses before they are aggregated into spatial-unit-level representations and uses them for prototype-based spatial relation learning. Specifically, BPR-Net represents each spatial unit through bin-resolved temporal responses, incorporates them into the prototype routing cost via Bin-Prototype Affinity, and modulates the routed representation using a spectral signature. Experiments on nine datasets across mobile traffic, air quality, taxi flow, and traffic sensor domains show that BPR-Net achieves the best result in 26 out of 28 dataset–metric settings against 19 baselines. Additional analyses show that BPR-Net achieves low forecasting error with short inference time in the reported efficiency settings and maintains low forecasting errors for spatial units with different spectral properties.

Accurate traffic forecasting is essential for Intelligent Transportation Systems (ITS). Attention-based Spatio-Temporal Graph Neural Networks (STGNNs) have been widely adopted to capture long-range spatial dependencies, but their node-level attention incurs quadratic complexity and limits scalability in large-scale traffic networks. Although sparse and hierarchical attention mechanisms reduce this cost by restricting node pairs, they still rely on node-level modeling. Through empirical analysis on a 2,352-node benchmark, we find that removing node-level attention from representative baselines does not consistently degrade performance, suggesting that its additional predictive benefit may be marginal in large-scale settings. In contrast, region-level attention appears useful for modeling traffic propagation, which is also consistent with classical traffic engineering models that represent traffic dynamics over aggregated spatial units. Motivated by these observations, we propose STAAR, a Spatio-Temporal Attention over Adaptive Regions framework that dynamically constructs soft traffic regions and models dependencies among them. Specifically, adaptive region assignment assigns each node to multiple regions with normalized weights, while spatio-temporal attention captures time-lagged dependencies by flattening region tokens across time steps into a unified sequence. Extensive experiments on large-scale traffic benchmarks demonstrate that our model achieves competitive accuracy with substantial efficiency gains, reducing training time by 3.5×–29× and inference time by 2.8×–28× compared with existing attention-based STGNNs. The source code of STAAR is publicly available at https://github.com/kaist-dmlab/STAAR.

Intelligent transportation systems (ITS) increasingly deploy roadside units (RSUs) equipped with sensors for environmental perception. Each RSU captures the surrounding traffic scene as a stream of point clouds, whose volume varies with vehicle density, mutual occlusion, and scene geometry. Since downstream tasks such as object detection, tracking, and cooperative perception depend on these streams, accurately predicting future perception volume is necessary for proactive resource provisioning in real-time ITS applications. However, this volume cannot be reliably determined from conventional traffic variables alone, as the same number of vehicles can produce substantially different amounts of point cloud data under different sensing conditions. Moreover, the perception influence between road sections shifts continuously as vehicles move and alter occlusion conditions, and vehicle arrivals introduce time-varying propagation delays that a fixed road-topology graph cannot capture. To address these challenges, we propose CPTIF, a continuous-time spatiotemporal forecasting framework that models each RSU as a node on a dynamic graph with adaptive edge weights fusing learnable structural priors and state-dependent dynamic relations. Temporal evolution and spatial propagation are unified through a cascaded neural controlled differential equation architecture, where discrete observations are first interpolated into continuous control paths and latent states evolve along these paths, enabling prediction at arbitrary future timestamps without retraining. We construct a perception traffic forecasting benchmark by coupling SUMO-driven traffic simulation with CARLA-rendered sensor perception across five topologically distinct maps. Experiments demonstrate that CPTIF outperforms representative baselines in prediction accuracy and multi-step stability.

\begin{abstract} Traffic flow forecasting requires modeling rapidly changing local fluctuations and slowly evolving regional trends. Existing Neural Ordinary Differential Equation (Neural ODE)-based traffic forecasting models usually evolve states at a single spatial granularity, which may entangle node-level volatility with macroscopic evolution and lead to over-smoothed peak predictions under volatile traffic conditions. To capture such mixed-scale dynamics, we propose Hierarchical Slow-Fast Spatial-Temporal Graph ODE (HSF-STGODE), a region-guided slow-fast continuous framework for traffic flow forecasting. HSF-STGODE decomposes traffic evolution into two coupled continuous processes: a Fast Graph ODE modeling node-level short-term fluctuations on the sensor graph, and a Slow Graph ODE capturing region-level trends on a learnable latent region graph. The two processes are connected through bottom-up aggregation and top-down guidance, enabling regional trends to provide stable context for fine-grained node evolution. To further model lagged propagation and irregular local variations, we incorporate a delay-aware pattern module, a flow-consistency regularizer, and a dual-gated fusion mechanism for scale-specific multi-period initialization. Experiments on four widely used PeMS benchmarks show HSF-STGODE achieves the best or tied-best RMSE across all datasets, with strong MAPE performance and competitive MAE results against representative and strong baselines. These results suggest region-guided slow-fast continuous modeling is particularly effective in reducing large forecasting deviations for traffic flow forecasting. Code is available at \url{https://github.com/KYasmaine/HSF-STGODE}. \end{abstract}

Traffic sensors commonly record flow, speed, and occupancy, but standard traffic flow forecasting benchmarks and models rarely exploit all three raw measurements reliably. Although speed and occupancy provide sensor-native traffic-state information beyond flow alone, existing releases often omit these variables, replace them with proxies, or contain logically inconsistent records. Moreover, direct empirical risk minimization over three-variable inputs may exploit regime-dependent shortcuts, as the relationships among flow, speed, and occupancy vary substantially between free-flow and congested states. We introduce \textbf{PEMSB-3V}, a public benchmark suite that preserves raw flow, speed, and occupancy measurements from PeMS detectors for flow prediction. We also propose \textbf{RiskTraf}, a model-agnostic risk-extrapolated residual plug-in. For each trained spatio-temporal backbone, RiskTraf freezes the selected checkpoint and learns a lightweight zero-start residual head from historical speed and occupancy. The residual head constructs ordered traffic-risk environments and optimizes horizon-wise flow corrections with a risk extrapolation objective, thereby mitigating regime-specific shortcut correlations without modifying the backbone. Extensive experiments demonstrate that RiskTraf consistently improves diverse forecasting backbones and outperforms debiasing and distribution-shift adaptation methods. Our code and benchmark are available at \url{https://github.com/Guangyu4/RiskTraf}.

Session-based Recommendation (SBR) aims to predict users' next interaction based on short-term anonymous behavior sequences. Yet accurately modeling the dynamic switching and hierarchical dependency of user intent remains a critical challenge. Existing methods typically treat intents as parallel signals thus ignoring the dynamic shift in the dominant decision-making pattern layer within a session. However, we argue that the dominant decision-making pattern shifts across hierarchical intent layers over time, and treating these layers as static or independent inevitably causes mismodeling. To address these issues, we propose a \textbf{P}yramid \textbf{R}outing for \textbf{I}ntent \textbf{S}equence \textbf{M}odeling framework (\textbf{PRISM}), which constructs an intent pyramid to parse user decision-making processes. Specifically, we first devise a Hierarchical Vector Quantization module to progressively compress actions into theme and goals. We then employ a Multi-Intent Routing Network to model the cross-layer dependency and information propagation among different intent layers. Finally, to adapt to the dynamic shifting of decision dominance, we design a Mixture of Intent Module that adaptively weights the prediction distributions from different layers, and introduce a Goal-Aware Contrastive Regularizer to ensure the consistency of hierarchical intent representations. Experimental results on three real-world datasets exhibit that PRISM achieves state-of-the-art performance.

Sequential recommender systems struggle to simultaneously capture persistent long-term preferences and volatile short-term intentions, while existing dual-stream architectures are computationally inefficient when handling long sequences and suffer from feature entanglement between the two interest streams. To address these issues, we propose a Mamba-based Adaptive Prototype LEarning and cross-intent contrastive alignment framework (MAPLE). MAPLE employs Mamba as the long-term encoder; its selective state space mechanism selectively retains important historical signals across the entire interaction sequence, making it well-suited for capturing stable long-term interest features with linear computational complexity. Meanwhile, the GRU-based short-term encoder focuses on a local tail window to model recent preference shifts. To decouple these two interest streams, we maintain a set of global semantic prototypes for each interest type and extract soft proxies as personalized semantic anchors through attention-weighted aggregation, thereby leveraging global intent consensus to enrich individual user representations. These proxies are further applied as contrastive anchors in a novel cross-intent InfoNCE loss, which treats proxies of the complementary intent stream as cross-intent contrastive anchors to promote semantic differentiation between long and short-term representations. Finally, we introduce an asymmetric dynamic gated fusion module that prioritizes long-term features as the main component and incorporates short-term features as the auxiliary component for next-item prediction. Experiments on four real-world benchmark datasets demonstrate that MAPLE achieves highly competitive performance, significantly outperforming state-of-the-art models on rank-aware metrics.

Sequential recommender systems commonly learn item representations from interaction data. This paradigm works well for frequent items, but becomes unreliable for long-tail items whose ID embeddings receive limited updates. Offline semantic embeddings derived from large language models provide useful item-side knowledge beyond interaction logs. They can therefore compensate for the insufficient interaction supervision of long-tail items. However, directly optimizing such semantic priors for recommendation may distort their original semantic structure. Moreover, semantic information should not be injected uniformly across all items, since head items are already well supported by interactions while long-tail items require semantic compensation more urgently. In this paper, we propose DART, a constrained semantic enhancement framework that couples Residual Trust Semantic Adaptation with Tail-Aware Dual-View Alignment for sequential recommendation. DART treats offline LLM-derived item embeddings as frozen semantic anchors and injects collaborative signals through a trust-region-constrained residual pathway. This design allows item representations to absorb useful collaborative information while limiting semantic drift. To further improve long-tail recommendation, DART introduces a tail-aware dual-view alignment mechanism that concentrates cross-view transfer on long-tail items instead of aligning all items uniformly. Experiments on three public benchmarks show that DART achieves strong overall performance and consistent gains in long-tail settings. Further ablation and representation analyses verify the effectiveness of residual-trust adaptation and tail-aware semantic transfer.

Sequential recommendation aims to predict users' future interests from their historical interactions. Although Large Language Models (LLMs) capture rich item semantics, existing methods often struggle to align collaborative signals with textual semantic knowledge. As a result, the learned item representations fail to capture the complementary strengths of both signals, leading to suboptimal recommendation quality. To address this limitation, we propose \textbf{L}atent \textbf{S}emantic \textbf{F}usion for \textbf{S}equential \textbf{R}ecommendation via Flow-based Conditional Variational Autoencoders (\textbf{LSF-SR}), a novel framework that uses a Conditional Variational Autoencoder (CVAE) with Normalizing Flows to fuse item ID embeddings and LLM-generated semantic signals. At the core of LSF-SR is a conditional fusion module augmented with planar or radial flows. This module learns a flexible latent space that encourages items with similar semantic profiles to cluster together within the latent manifold. Through extensive experiments on five public benchmark datasets, we demonstrate that LSF-SR consistently outperforms state-of-the-art baselines, achieving gains of up to 12.98\% and 14.13\% in Recall@20 and NDCG@20, respectively.

Sequential recommendation aims to model users’ evolving preferences from their historical interactions. Recent latent reasoning methods improve sequential recommenders by extending inferencetime computation in continuous representation space, refining user representations without generating explicit reasoning text. However, existing methods mainly reason over the encoded history itself and make limited use of contextual feedback from users’ recent interactions. Such feedback, especially positive and negative signals in the latest session, can not only complement the long-term user representation but also provide directional guidance for latent reasoning by indicating which preference patterns should be strengthened or suppressed. In this paper, we propose Context-Aware Latent Reasoning (CaLR), a model-agnostic framework that grounds latent reasoning in session-level context. From the input perspective, CaLR selects informative positive and negative items from users’ recent sessions as reasoning anchors. From the architecture perspective, it injects these anchors into each reasoning step through a Cross-Attention-based module. From the optimization perspective, it introduces Process Supervision Loss (PSL) to supervise intermediate reasoning states according to anchor feedback polarity. Experiments on two real-world public benchmarks show that CaLR consistently improves multiple sequential recommendation backbones and outperforms strong latent reasoning baselines. Ablation and further analyses verify the effectiveness of each component in CaLR. Moreover, a seven-day online A/B test on Kuaishou Lite demonstrates its effectiveness in real-world industrial deployment. Our code is available at https://github.com/Ak4ne-windsky/CaLR.

Sequential recommendation aims to predict the next item a user will interact with from behavioral history. Frequency-domain methods emerge as a promising approach to capture different behavioral patterns by decomposing user sequences into frequency components with various filters. However, existing spectral filters mainly emphasize or suppress frequency components, without explicitly modeling how preference signals evolve in the spectral domain. To overcome this limitation, we propose \textbf{TelaRec}, a novel physics-inspired sequential recommendation model built upon the Thermoelastic Spectral Operator (TSO). TSO treats user behavior sequences as preference fields and models their evolution through heat conduction and elastic wave propagation, where heat conduction extracts stable preferences and elastic wave propagation captures short-term interests. Derived from partial differential equations, TSO enables analytical spectral operations beyond conventional static or adaptive filtering. We further introduce User-aware Frequency Calibration (UFC) to adjust the propagation scale of each frequency mode according to the current user's spectral response for personalized spectral evolution. Experiments on six benchmark datasets demonstrate that TelaRec achieves competitive recommendation performance and high computational efficiency. Our source code is available at https://github.com/TBI805/TelaRec.

Sequential recommenders are typically trained on long user histories to capture rich behavioral signals, yet serving with training-length sequences is often impractical due to real-time efficiency constraints. Directly using only recent behaviors leads to a severe performance drop. To bridge this gap, existing approaches compress user histories into persistent per-user states, storing and retrieving them at inference time; while effective, they impose non-trivial infrastructure overhead and offer little remedy in cold-start scenarios. In this paper, we empirically identify two structural flaws rooted in geometric properties and dataset sparsity, and propose a novel two-stage framework to close the long-short-view performance gap. Specifically, in the first stage, we replace the commonly used dot-product with angular similarity scoring and leverage a modified softmax to counter prefix position bias. In the second stage, we fine-tune only bias and LayerNorm components—universal to standard sequential backbones—for further improvement. Both stages are guided by carefully designed learning objectives. Extensive experiments on two representative backbones across three public datasets demonstrate the effectiveness of our proposed framework.

Sequential recommenders often struggle with sparse and noisy histories, limiting generalization to unseen interactions. Knowledge distillation mitigates this by transferring dense supervision from a teacher to a student. However, most distillation methods run teacher and student independently, then match student outputs or representations to the teacher. Such supervision entangles student-component effects, blurring whether weak generalization stems from unreliable embeddings, overfitted encoding, or co-adaptation to sparse histories. In this paper, we propose \textbf{G}raft-\textbf{O}riented \textbf{D}istillation (\textbf{GOD}), a component-level distillation framework for improved generalization through grafting. Grafting denotes replacing selected frozen-teacher components with trainable student counterparts to build hybrid source models. GOD uses these hybrid models to evaluate student embeddings with the teacher encoder and the student encoder with teacher embeddings, providing component-level feedback. At inference, GOD uses only the student, incurring no additional cost. Across three real-world datasets, GOD outperforms state-of-the-art baselines by up to 13.92\%.

Re-ranking aims to present users with an ordered list that maximizes sequence-level utility. Industrial deployments widely adopt the Generator--Evaluator framework, where the generator proposes candidate sequences, and the evaluator scores them to select the best one for display. The fidelity of the evaluator's utility estimates thus directly governs recommendation quality. However, existing evaluators predominantly follow the \textbf{Independent Behavior Modeling (IBM)} paradigm, which simplifies the continuous user \emph{browsing journey} into independent per-item predictions (e.g., CTR). This oversimplification ignores two physical realities: (i) a click can occur only after the item has been reached and impressed, and (ii) items compete for finite user attention. As a result, IBM produces systematically biased evaluations and suboptimal sequence selection. To address this, we propose a \textbf{Joint Behavior Modeling (JBM)} perspective that explicitly models the \emph{display} $\rightarrow$ \emph{impression} $\rightarrow$ \emph{click} process to capture causal and competitive dependencies within a sequence. Building on JBM, we develop the \textbf{Impression-Click Evaluator (ICE)}, a plug-and-play framework that (i) predicts a browsing-depth distribution to estimate per-position impression probabilities, (ii) leverages predicted impressions to suppress spurious context from unexposed items during click modeling, and (iii) introduces a click distribution matching objective to capture intra-sequence dependencies. Extensive offline experiments across diverse backbones, together with a 14-day online A/B test in the Taobao App (+0.70\% clicks, +3.67\% revenue) and full-traffic deployment, validate the effectiveness and robustness of our approach.

Large-scale live-streaming recommendation requires precise modeling of non-stationary content semantics under strict real-time serving constraints. In industrial deployment, two common approaches exhibit fundamental limitations: discrete semantic abstractions sacrifice descriptive precision through clustering, while dense multimodal embeddings are extracted independently and remain weakly aligned with ranking optimization, limiting fine-grained content-aware ranking. To address these limitations, we propose \textbf{SARM}, an end-to-end ranking architecture that integrates natural-language semantic anchors directly into ranking optimization, enabling fine-grained author representations conditioned on multimodal content. Each semantic anchor is represented as learnable text tokens jointly optimized with ranking features, allowing the model to adapt content descriptions to ranking objectives. A lightweight dual-token gated design captures domain-specific live-streaming semantics, while an asymmetric deployment strategy preserves low-latency online training and serving. Extensive offline evaluation and large-scale A/B tests show consistent improvements over production baselines. SARM has been fully deployed and serves over 400 million users daily at Kuaishou.

Stale recommendations are a pervasive challenge and a leading source of user complaints on large-scale content platforms. Items lose relevance through two primary mechanisms: supersession, where emerging updates render prior coverage stale, and relevance decay, where an item's informational value naturally diminishes over its lifecycle. Traditional countermeasures serve as crude proxies: age cutoffs poorly reflect actual relevance loss, while engagement heuristics rely on lagging signals, broadly exposing users to stale content before the system adapts. We present SDF (Supersession-Decay Filtering), a staleness filtering system fully deployed in Google Discover, a personalized recommendation feed with hundreds of millions of daily and billions of monthly active users. SDF targets both mechanisms with complementary filters, each powered by a learned model: a relational staleness model that detects supersession between item pairs, and a predicted traffic ratio (PTR) model that forecasts relevance decay from the item's content, trained on lifetime visit traffic. Applied via disjunction upstream of the ranking stage, SDF prunes stale candidates, measurably reducing downstream serving costs. Online experiments demonstrate that these filters significantly reduce the prevalence of stale content while improving user engagement. Over a two-year production deployment, user-filed staleness reports (in-product user feedback) declined by 54.9% relative to the pre-deployment baseline, establishing SDF as a robust and scalable paradigm for resolving content staleness at industrial scale.

Temporal distribution shift (TDS) degrades the generalization of recommender systems: models trained on past interactions face systematically different user behavior in future time windows. While incremental training on recent data is widely used in industry production, it can over-adapt to transient patterns and exacerbate instability. Existing approaches such as invariant learning and self-supervised learning offer partial solutions but often suffer from unstable temporal generalization, representation collapse, or inefficient data utilization. To address these limitations, we propose ELBO$_\text{TDS}$, a probabilistic framework for temporally robust recommendation. First, we identify key shifting factors on real-world production data and design a simple yet effective data augmentation strategy that simulate realistic time-varying factors to extend the training support. Second, we derive a self-supervised variational objective ELBO$_\text{TDS}$ to harness the benefits of this extended distribution while preventing representation collapse. Extensive experiments on both a public benchmark and a large-scale industrial dataset demonstrate ELBO$_\text{TDS}$ consistently improves next-period ranking performance over strong baselines, yielding a 2.33\% uplift in GMV per user online and has been successfully deployed in production. Code is available at \zhu{\url{https://github.com/FuCongResearchSquad/ELBO4TDS}}.

At Amap, a leading mapping platform serving hundreds of millions of users, isolated search and recommendation systems posed significant challenges in engineering costs and user experience. To address this, we present IntSR, a generative system that successfully integrates search and recommendation (S\&R) tasks. IntSR has delivered substantial business value across diverse scenarios, including a +9.34\% relative increase in digital asset GMV, a +2.76\% relative lift in POI recommendation CTR, and a +7.04\% relative improvement in travel mode suggestion ACC. The system's success relies on two innovations designed for a real-world environment: (1) a novel Query-Driven Block (QDB) to manage the computational complexity of a unified model at scale, and (2) a time-aware negative sampling strategy to handle constantly changing corpus items. Our work provides a valuable reference and critical insights into deploying integrated generative models in industrial settings.

Applying scaling laws to recommendation retrieval is hindered by feature heterogeneity: naively stacking Transformer layers yields diminishing returns because heterogeneous fields produce severe token-norm divergence. We present \textbf{TransRetrieval}, a Transformer-based retrieval framework that scales with both computational budget and cross-domain data. The key enabler is (1)~weighted average aggregation, which restores the homogeneous-token assumption Transformers rely on. Building on this, we introduce (2)~target token compression that cuts per-candidate FLOPs by 85\% while preserving cross-attention expressiveness, and (3)~position-style domain embeddings that unify multiple domains at negligible additional cost, turning cross-domain data into a scaling asset. On a 40-billion-interaction industrial dataset and the public KuaiRand benchmark, scaling compute from 0.1 to 2\,MFLOPs per target yields +19.3\,/\,+22.2\,pt Recall@2000, confirming robust log-linear scaling. In online A/B tests, TransRetrieval lifts platform revenue by 2.53\% under the same end-to-end latency constraint as the production baseline.

While Transformers have achieved remarkable success in LLMs through superior scalability, their application in industrial-scale ranking models remains nascent, hindered by the challenges of high feature sparsity and low label density. In this paper, we propose SORT (\textbf{S}ystematically \textbf{O}ptimized \textbf{R}anking \textbf{T}ransformer), a scalable model designed to bridge the gap between Transformers and industrial-scale ranking models. We address the high feature sparsity and low label density challenges through a series of optimizations, including request-centric sample organization, local attention, query pruning and generative pre-training. Furthermore, we introduce a suite of refinements to the tokenization, multi-head attention (MHA), and feed-forward network (FFN) modules, which collectively stabilize the training process and enlarge the model capacity. To maximize hardware efficiency, we optimize our training system to elevate the model FLOPs utilization (MFU) to 45\%. Extensive experiments demonstrate that SORT outperforms strong baselines and exhibits excellent scalability across data size, model size and sequence length, while remaining flexible at integrating diverse features. Finally, online A/B testing in large-scale e-commerce scenarios confirms that SORT achieves significant gains in key business metrics—including orders (+7.47\%), buyers (+6.67\%) and GMV (+8.65\%)—while simultaneously cutting latency by 62\% and boosting throughput nearly sevenfold (+589\%). SORT has been fully deployed in production, serving all users on AliExpress.

Scaling up recommender models has become a major route to improving recommendation performance, but effective scaling relies on learning robust and discriminative representations. We identify \emph{local representation collapse} as a critical bottleneck in this process: under sparse feedback, label-distinct samples can be mapped to highly overlapping local neighborhoods, making fine-grained discrimination increasingly difficult. Existing studies of representation collapse mainly focus on global representation statistics, while auxiliary ranking objectives typically apply pairwise supervision broadly across samples, overlooking the localized nature and varying severity of collapse. To address this limitation, we propose \emph{Auxiliary Focal Discrimination (AFD)}, a geometry-aware framework that adaptively concentrates discriminative supervision on locally collapsed label-distinct pairs. AFD measures pairwise latent proximity using a closeness coefficient and employs it to weight an auxiliary ranking loss, assigning stronger supervision to more severely collapsed pairs while exerting less influence on already well-separated ones. Experiments on public benchmarks demonstrate that AFD consistently improves ranking performance, enlarges positive--negative representation margins, and slows singular-value decay, indicating stronger discriminability and alleviated dimensional collapse. Moreover, AFD has been fully deployed in Tencent's online advertising system, achieving GMV lifts of $0.80\%$ and $1.06\%$ in Weixin Channels and Weixin Moments, respectively.

Recent work leverages large-scale pretrained foundation models (FMs) to transfer generalizable representations to downstream advertising tasks. Most existing paradigms transfer user representations (URs) and item representations (IRs), while underutilizing the user-item cross representations (CRs) generated by deeper interaction layers. Although a recent approach transfers target-aware embeddings (TAEs), it requires real-time upstream FM inference alongside downstream model inference, introducing additional engineering complexity and overhead. To address these limitations, we propose Large Foundation Model for Ads Recommendation (LFM4Ads), an efficient \textit{pretrain-then-transfer} framework for cross-domain ad recommendation. LFM4Ads jointly exploits URs, IRs, and CRs from the upstream FM, covering both isolated user/item semantics and fine-grained user-item interactions. It extracts sample-level CRs from pretrained interaction layers and aggregates them into user-level CRs with temporal decay, thereby decoupling upstream FM inference from downstream online serving. To enable efficient utilization of CRs, we further propose \emph{multi-granular downstream usages}, including feature-level gating for CR fusion, module-level fine-tuning to yield downstream CRs, and model-level retrieval by matching CRs with IRs. Since Q4 2024, LFM4Ads has been deployed on Tencent's advertising platform. It has supported 10+ launches across major scenarios, yielding a platform-wide gross merchandise value (GMV) lift of 2.45\% and annual revenue gains in the hundreds of millions of dollars. Offline experiments on KuaiRand-1K also validate the effectiveness of LFM4Ads, and our code is available at \url{https://github.com/ZhuangQu/LFM4Ads}.

Industrial ranking models for recommendation have scaled feature interaction and sequence modeling separately; recent architectures such as HyFormer and MixFormer unify both in a stackable backbone. Real-world recommender systems, however, nearly always require multi-task learning, yet existing unified architectures confine multi-task modeling to shallow post-backbone towers, leaving the backbone without task-aware information flow. We propose MORE (Multi-task cO-evolving Ranking modEl), which embeds multi-task information flow inside the backbone, enabling task-specific signals to co-evolve with sequence and feature representations at every layer rather than in a post-hoc fusion. It introduces Anchor Tokens that persist across backbone layers: Shared Anchors encode cross-task commonalities, while Private Anchors capture task-specific priors. In each block, Anchor Tokens (1) read task-conditioned signals from behavior sequences, (2) mix with non-sequential features under a task-boundary mask, and (3) refine per-task representations through independent branches; as blocks stack, each task obtains a differentiated representation refined through all backbone layers. Experiments on large-scale industrial datasets show that MORE consistently outperforms baselines across all tasks under comparable parameter and FLOPs budgets, and scales well with model size. Online A/B tests on Momo, a leading Chinese social discovery platform with tens of millions of monthly active users, yield 3% improvement in usage duration, 3.6% in interaction rate, and 2% in deep-chat rate. MORE is deployed in production with request-level shared computation reducing scoring latency by about 30%.

LinkedIn Feed enables professionals worldwide to discover relevant content, build connections, and share knowledge at scale. We present Feed Sequential Recommender (Feed SR), a transformer-based sequential ranking model for LinkedIn Feed that replaces a DCNv2-based ranker and meets strict production constraints. We detail the modeling choices, training techniques, and serving optimizations that enable deployment at a scale of 1.2 billion members. Feed SR has been serving the majority of LinkedIn’s Feed traffic for over three months and shows significant improvements in member engagement (+2.10% time spent, +3.52% like, comments, or reshares) in online A/B tests compared to the existing production model. We also describe our deployment experience with alternative sequential and LLM-based ranking architectures and why Feed SR provided the best combination of online metrics and production efficiency.

Friend recommendation is inherently graph-structured: the relevance of a potential connection depends on multi-hop social context rather than user attributes alone. However, deploying message-passing GNNs on a production-scale social graph with hundreds of millions of users and tens of billions of edges requires addressing numerous modeling and systems challenges. We present a scalable end-to-end GNN ranking system for production social graphs, focusing on two design choices that are critical in this setting: multi-hash ID embeddings and temporal neighbor sampling. Multi-hash embeddings are common for high-cardinality features, but industrial GNN systems typically either ignore trainable IDs or accept full embedding tables — exceeding 200 GB for our graph. We integrate multi-hash as the primary node representation, reducing the ID-embedding table size by >98% while preserving ranking quality. Temporal neighbor sampling is well understood in principle, but existing implementations scan full adjacency lists — a non-starter for users with tens of thousands of friends. We implement timestamp-sorted CSR storage with binary search, reducing the per-node temporal sampling cost from O(deg(v)+k) to O(log deg(v)+k). Beyond these components, we show that this combination scales and yields measurable production impact. On a graph with 194M users and 28B edges, offline ablations isolate each design choice's contribution. In an online A/B test, our system increases friend additions from recommendations by 16% and unique friend adders by 11.5% over a strong production baseline. We release our framework for distributed training and inference on large temporal graphs.

Large Language Models (LLMs) are increasingly adopted for compliance and legal reasoning tasks, yet their outputs often lack explicit grounding in legal logic and evidence. We present Code-as-Auditor, an LLM-based framework that extends the model’s reasoning capability toward structured and evidence-grounded compliance assessment. The framework translates regulatory information into (1) formalized checklists and executable decision trees, encoding regulations and conditions as interpretable code structures. During inference, each checklist item is (2) dynamically expanded into factual and counterfactual questions, guiding the model to reason over case-specific evidence and potential violations. This process establishes a reasoning pipeline that proceeds from evidence identification, through rule application, to final decision-making, while a self-verification loop ensures logical consistency and traceability of outcomes. Experiments on privacy and data protection scenarios demonstrate that Code-as-Auditor delivers more accurate and evidence-backed evaluations, enabling automated compliance regulation checking grounded in explicit regulatory criteria.

Narrative reasoning is defined as predicting future events from narrative context, requiring an understanding of event evolution and implicit causality. Current memory-augmented approaches follow two separate paradigms: episodic-memory retrieval, which leverages past analogies but often leads to logical hallucination due to surface-level similarity, and semantic-memory retrieval using Event-centric Knowledge Graphs (EKGs), which ensures logical consistency but suffers from sparsity in open-world settings. Inspired by the Complementary Learning Systems (CLS) theory in cognitive science, we propose a Dual-Memory Neuro-Symbolic Framework (DMNeS), which synergizes episodic intuition and semantic verification for logically reliable narrative reasoning. Specifically, the Episodic Intuition Reasoner generates intuitive predictions based on retrieved analogous episodes. To mitigate potential logical hallucination, these predictions are then validated by the Semantic Logic Verifier using causal-temporal evidence retrieved from an EKG. When an otherwise consistent candidate lacks an explicit EKG path due to sparsity, Abductive Reflection prompts the LLM to hypothesize a latent bridge event and connecting relations, and re-verifies the resulting chain under occurrence and temporal constraints. Finally, the Dynamic Memory Evolution module consolidates both memories, strengthening their interplay. As a fine-tuning-free inference framework, DMNeS achieves a 5.7 percentage-point average gain over the strongest baseline and improves LLM backbones by 7.5--9.7 percentage points across three benchmarks.

The discovery of first integrals is of profound scientific significance for understanding physical conservation laws in dynamical systems. However, existing symbolic computation systems and Large Language Models (LLMs) struggle with this task due to the scarcity of high-quality training data and the reliance on mathematical intuition for finding solutions. This paper introduces \textbf{FISolver}, an LLM-based solver designed to bridge this gap. First, we propose a novel ``Backward Generation'' algorithm that systematically constructs large-scale datasets of (differential equations, first integrals) pairs by deriving equations from sampled integrals, effectively overcoming the data scarcity bottleneck. Second, we perform supervised fine-tuning on a compact mathematical model and further enhance its capabilities using reinforcement learning with a Levenshtein Distance-based shaped reward function. Additionally, we introduce data synthesis and blending techniques that enable effective adaptation to specific hard instances with sparse examples. Experiments demonstrate that \textbf{FISolver}, with significantly lower computational cost, substantially outperforms larger mathematical LLMs and commercial solvers like Mathematica on complex benchmarks, establishing a new data-driven paradigm for automated mathematical discovery.

Large language models increasingly shape human decision-making as conversational assistants and act on human behalf as autonomous agents, making the cultural values they encode consequential across both settings. Yet ensuring that these systems reflect the values of diverse cultural groups remains an open challenge, as existing alignment approaches offer limited ability to precisely control cultural value orientations on demand. In this work, we investigate cultural value representations in LLM activation space through the lens of Hofstede's six-dimensional framework. We first demonstrate that contrastive activation steering reliably controls individual cultural value dimensions, significantly and consistently outperforming prompt-based steering across two models and two independent evaluation protocols. We then show that cultural value dimensions are not geometrically independent: positively correlated value vectors mutually amplify each other's effects at the cost of response coherence, while negatively correlated vectors attenuate each other through directional cancellation. These interference patterns impose a fundamental constraint on naive multi-dimensional steering. Motivated by this geometric analysis, we propose a selective projection decoupling method that removes inter-dimension interference prior to joint injection and fits steering coefficients to country-level Hofstede profiles. Evaluated on Brazil, China, and the United States, our method achieves consistent improvements in alignment accuracy, direction hits, and cultural similarity across both Qwen-2.5-7B-Instruct and Ministral-8B-Instruct models, demonstrating that steered models are not only more culturally accurate but achieve high cultural similarity to their intended target profiles.

Large language models (LLMs), especially recent reasoning-oriented models trained with reinforcement learning, achieve near-saturating accuracy on many reasoning benchmarks. However, it remains unclear whether these models truly reason over semantic structure or rely on superficial symbol patterns and familiar language forms. This issue is increasingly critical in the agent era, where reasoning systems must remain reliable under changing representations. We introduce FLoRE, a controllable evaluation benchmark for diagnosing LLM reasoning robustness and effectiveness using formal languages. By isolating reasoning from world knowledge, formal languages enable direct evaluation of structural reasoning with machine-verifiable ground truth. We defines two tasks: Formal Language Recognition and Grammar Induction and construct 25.5k data instances. Evaluating diverse reasoning and non-reasoning LLMs, we identify two universal failure modes. First, models exhibit severe symbol-binding fragility: changing symbolic representations while preserving semantics causes large drops on positive instances but increases accuracy on negative instances, revealing a strong default rejection bias under unfamiliar symbols. Second, many reasoning-oriented local LLMs show negation non-termination during grammar induction, failing to terminate reasoning on negative instances and instead exhausting large portions of the context window. Our results show that current LLM reasoning remains strongly dependent on surface symbol priors and lacks robust semantic invariance and termination stability. FLoRE provides a unified framework for systematically evaluating these properties in controllable formal-language settings. Code and data files are available at \url{https://github.com/HoganZinger/FLoRE}

As Large Language Models (LLMs) are evaluated on increasingly large and diverse benchmark suites, running an LLM on the entire benchmark incurs substantial costs. We study efficient LLM evaluation, where the goal is to accurately estimate full-benchmark performance by adaptively querying only a small subset of prompts. The key challenge is that selected prompts should be informative for the target model while remaining representative of diverse response patterns in the benchmark. To address this, we adopt a posterior-aware Bayesian perspective: historical response data are used to characterize complex prompt--model behaviors, and the resulting posterior information guides adaptive prompt selection under a limited budget. Specifically, we develop a variational Gaussian Process Item Response Theory (VGP-IRT) model, which leverages item-specific Gaussian process response functions to flexibly capture nonlinear prompt--model interactions beyond restrictive parametric response curves. Based on prompt features and predictive information derived from the VGP-IRT posterior, we further construct group-balanced candidate pools within local routed subspaces for adaptive selection. Experiments on four benchmarks demonstrate improved score estimation through our method, with the strongest gains in the low-to-moderate budget regime.

Vehicle diagnosis begins with structured signals---an Electronic Control Unit (ECU) identifier, a Diagnostic Trouble Code (DTC), and trigger conditions---but technicians must translate them into an accurate fault description and appropriate repair measures. Automotive repair reasoning is both long-tailed and safety-critical: rare ECU/DTC combinations provide few training examples, while associating a fault with the wrong control unit or recommending unnecessary repairs can mislead maintenance decisions. Benchmarking this capability is challenging because diagnostic resources are fragmented, general-purpose large language models (LLMs) often fail to follow repair constraints, and exact-match metrics may reject valid answers expressed differently. We introduce \textbf{AutoRepairBench}, a BMW-centered benchmark and evaluation framework for long-tailed, safety-critical automotive repair reasoning. It supports single-turn diagnosis, multi-pair prompting, and frequency-aware analysis of common and rare cases. Its validated hierarchical evaluator accepts valid paraphrases while checking ECU/DTC associations, component relationships, repair severity, and unnecessary repair recommendations. We also develop \textbf{RepairLLM-7B}, a compact, retrieval-free model trained through two-stage distillation: the first stage reduces the dominance of frequent cases, while knowledge-graph-anchored online distillation emphasizes uncertain and rare cases. On a double-blind set of 585 cases labeled by three automotive experts, the evaluator achieved 94.19\% accuracy and an F1 score of 94.67\%. On the 500-instance main evaluation set, RepairLLM-7B achieved 95.0\% single-turn joint accuracy (475/500) and 94.2\% per-pair joint accuracy under multi-pair prompting (471/500); it also improved tail accuracy on the separate 100-instance generalization set and reduced safety-critical errors relative to supervised fine-tuning. These results demonstrate the importance of jointly benchmarking overall accuracy, rare-case performance, and repair safety in automotive repair reasoning.

Large vision-language models (LVLMs) are increasingly used for multimodal search, open-ended question answering, document understanding, and generated content analysis, yet they still produce claims that are fluent but unsupported by visual evidence. Existing hallucination mitigation methods often rely on external multimodal judges, language-model annotators, or human token labels to localize unsupported content, adding cost and introducing a mismatch with the target model's own failure modes. We propose HIS-Guard, a framework for trustworthy multimodal claim generation that diagnoses unsupported claims using the target LVLM's internal stability signals and routes those signals to selective verification or preference optimization. Its core diagnostic, Hallucination Instability Score (HIS), combines semantic convergence from cross-layer LogitLens entropy reduction with visual grounding coherence from cross-token attention-distribution consistency over image patches. By decomposing HIS into semantic and visual instability, HIS-Guard supports claim-level credibility analysis without external token-level hallucination localizers or auxiliary multimodal judges. We then instantiate a Diagnose-Treat-Prevent preference objective (DTP-DPO): semantic instability treats risky tokens by weighting DPO rewards, while visual and combined instability prevent visual neglect through gated auxiliary regularizers. Across Qwen2.5-VL-3B/7B and InternVL2.5-2B, HIS-guided optimization provides downstream evidence that the diagnostic captures useful failure modes. On Qwen2.5-VL-3B, HIS-Guard reduces CHAIR_s by 32.2% and CHAIR_i by 45.6% over the baseline, improves over standard DPO trained on identical data, and largely preserves POPE accuracy. Scaling to Qwen2.5-VL-7B keeps HIS-Guard best on both CHAIR and POPE, while InternVL results suggest that the diagnostic signal is useful but the regularization strength is architecture-sensitive. Together, these results position internal stability as a practical signal for trustworthy multimodal information systems, selective verification, and hallucination-aware evaluation. The public implementation is available at https://github.com/20508Z/HIS-DTP-DPO.

In intelligent education systems, cognitive diagnosis models (CDMs) assess students' latent abilities through response records to enable personalized learning. However, existing CDMs struggle with inaccurate ability estimation under student cold-start conditions. Text-based CDMs, despite their strong generalization in data-rich scenarios, still suffer from ambiguous representations due to insufficient response records, which significantly hinders their practical deployment in real-world educational systems. Inspired by teachers' analogical reasoning when assessing new students, we propose Retrieval-Augmented Cognitive Diagnosis (RACD), a retrieval-augmented framework for CDMs. RACD fundamentally addresses student cold-start problems through a memory-retrieval-aggregation module that emulates teachers' analogical reasoning by retrieving and integrating similar student profiles from historical data. RACD is further enhanced by (1) Distinctiveness Regularization to align encoder outputs with distinctive representations and (2) Skill Shake that enforces contrastive learning for the retrieval module. Evaluations on three real-world datasets across four distinct cold-start configurations (different retention rates) show RACD outperforms both state-of-the-art CDMs and emerging LLM-based variants, with an average accuracy improvement of \textbf{1.2\%} and a peak gain of \textbf{2.1\%} over the best baseline in inductive scenarios, especially under student cold-start scenarios.

Understanding how students with different proficiency levels respond to educational materials is a critical issue within the field of Education. However, acquiring sufficient real student response data for a robust evaluation is often hindered by cost, ethics, and security constraints. Consequently, LLM-based student proficiency simulation, especially prompt-based methods, has emerged as a practical alternative under data-scarce conditions. Despite their promise, current methods still exhibit limited controllability with coarse-grained proficiency representations, high sensitivity to prompt design, and the lack of calibration with academic performance. Therefore, we propose \textbf{P}arameterized \textbf{S}tudent \textbf{P}roficiency \textbf{S}imulation (\textbf{\model}), an unsupervised and parameterized model-level framework that simulates students with different proficiencies by interpolating between a strong upper-bound LLM and a weaker, cognitive error-informed lower-bound student LLM via a hybrid ratio. Specifically, the lower-bound model is constructed by fine-tuning the weaker LM to exhibit cognitive errors when responding to educational materials. To ensure alignment with target proficiency levels, \model further calibrates the interpolation ratio with academic performance. Experiments on two public datasets demonstrate that \model achieves finer-grained and consistent proficiency simulation compared to existing baselines, leading to superior performance in student behavior similarity and item difficulty prediction.

Knowledge tracing (KT) aims to predict learners' future performance by modeling their historical interaction data. In real-world educational scenarios, multiple-choice questions are deliberately designed so that each distractor exposes a specific misconception; however, existing KT models rely on randomly initialized ID embeddings and reduce each interaction to a binary correctness label, making it difficult to exploit such option-level diagnostic signals. To address these problems, we propose an option-level pedagogical explanation framework for augmented KT, i.e., OPERA, which combines LLM-based option-level explanation generation, multi-judge quality assurance, and offline semantic embedding injection into existing KT backbones, with all LLM computation confined to one-time offline preprocessing. Experiments on three multiple-choice datasets demonstrate our framework's superior performance in terms of prediction accuracy, consistently improving lightweight ID-based backbones to match or surpass much heavier process-aware models. To encourage reproducible research, we have made our code publicly available at \url{https://pykt.org/}.

Knowledge Tracing (KT), a core component of intelligent tutoring systems, aims to model the dynamic evolution of students' knowledge states based on their historical interaction data. However, mainstream methods rely heavily on large-scale students' response logs, hindering generalization to unseen students, and often overlook the semantic structure of knowledge points in knowledge graph construction. In this paper, we propose Semantic Knowledge Graph-Based Knowledge Tracing (\textbf{SKG-KT}), a novel KT framework that leverages the reasoning ability of large language models (LLMs) to enhance the modelling of unseen students, without relying on numerous students with historical records. SKG-KT consists of two stages: ATC-ruled Semantic Graph Extraction and a Concept-aware Reasoning mechanism. In the first stage, we design a semantic student knowledge graph extraction method grounded in ATC rules, which captures the semantic structure across three hierarchical layers—domain, cluster, and standard—aligned with established educational theory. To extend beyond mathematics, we further introduce a multi-disciplinary semantic graph extraction method that induces an ATC-inspired hierarchical schema for other disciplines via LLM-based label induction. In the second stage, the Concept-aware Reasoning mechanism first applies knowledge concept reasoning to infer key knowledge concepts for unseen students with limited previous student data, and then fine-tunes an open-source LLM for correctness prediction based on the reasoned knowledge concepts, adapting the LLM to the KT task. Extensive experiments conducted on three publicly available educational benchmark datasets, namely Xes3g5m-small, Xes3g5m-large, and MOOCRadar, demonstrate that SKG-KT achieves state-of-the-art performance in knowledge tracing. Furthermore, SKG-KT has been validated to achieve robust predictive performance compared to others, while relying on a limited number of students with historical records. Our code and prompts are publicly available at the link \footnote{https://github.com/linqingliDestinyController/SKG\_KT/tree/main}.

Knowledge Tracing (KT) aims to assess students' dynamic knowledge states from their learning histories. While most existing KT methods focus on single-domain learning with notable success, real-world learning scenarios often involve multiple domains simultaneously, introducing two critical factors: 1) Cognitive load, arising from managing learning across domains in both temporal and knowledge dimensions. 2) Knowledge transfer, where knowledge states in one domain influence related states both within and across domains. In this paper, we focus on exploring these factors to improve students’ knowledge state assessment in multi-domain learning scenarios and propose a novel method incorporating cognitive \textbf{L}oad and knowledge \textbf{T}ransfer for \textbf{M}ulti-domain \textbf{K}nowledge \textbf{T}racing (\textbf{LT-MKT}). Specifically, to bridge isolated domains, LT-MKT first integrates textual information from questions and their associated concepts to construct a Multi-domain Hierarchical Graph, leveraging the advanced representational capabilities of large language models (LLMs). Then, cross-domain features in both the temporal and knowledge dimensions are explicitly modeled to capture the effects of cognitive load. Additionally, a knowledge transfer module is designed to model the propagation of knowledge states within and across domains. By jointly modeling these factors, LT-MKT enables more accurate prediction of students’ future performance. Finally, extensive experiments on real-world datasets demonstrate that our method achieves state-of-the-art performance.

Molecular property prediction is important for drug discovery, but labeled molecules are often limited for new prediction tasks. The difficulty is amplified in cross-domain few-shot settings, where target tasks differ from available source tasks and the model must adapt from sparse target-domain supervision. Existing few-shot and transferable few-shot methods mainly condition adaptation on the target support set, which can be unstable when the support molecules provide limited coverage of the target distribution. We propose Foundation-Assisted Cross-Domain Meta-Learning for Few-Shot Molecular Property Prediction (FACML), which augments cross-domain episodic learning with property-aware priors from a frozen Structure-Property Multi-Modal foundation model (SPMM). FACML represents each molecule using SPMM-derived SMILES features and graph features, and constructs support Property Vector (PV) prototypes and query PV representations from SPMM-predicted PVs based on molecular SMILES. A property-conditioned relation graph then integrates molecular nodes and PV-based prior nodes to refine episode representations for target-task adaptation. Across cross-dataset few-shot molecular property prediction benchmarks, FACML improves the average ROC-AUC over the strongest baseline by 6.36\% and 8.12\% under the 1-shot and 10-shot settings, respectively. An auxiliary zero-target-label analysis further examines the use of SPMM-generated molecules when labeled target support molecules are unavailable.

Medication recommendation requires both therapeutic appropriateness and control of drug-drug interaction (DDI) risk. However, existing approaches face a semantic mismatch between the molecular mechanisms underlying DDIs and the clinical abstractions used for therapeutic assessment. This work addresses this limitation through a dual-level formulation that models interaction risk at the molecular level while incorporating therapeutic supervision and evaluation at the clinical class level. Building on this formulation, we propose \textbf{$\mathrm{IM}^3$}, an interaction region-augmented multimodal molecular modeling framework that integrates longitudinal electronic health records with quantum-informed drug representations. $\mathrm{IM}^3$ employs the Interaction Region Indicator (IRI) to encode interaction-relevant electronic regions of drug compounds and combines global electronic characteristics with local chemical substructures. This representation incorporates molecular interaction information beyond static molecular topology and prescription co-occurrence. Experiments on two real-world intensive care unit datasets show that $\mathrm{IM}^3$ substantially reduces DDI rates while maintaining competitive therapeutic performance relative to state-of-the-art baselines. These results indicate the complementary value of dual-level modeling and quantum-informed molecular representations for safer medication recommendation.

Knowledge-grounded assessment of procedural compliance is a growing challenge in healthcare informatics. Such assessment benefits from explicit grounding in authoritative clinical references. Hand hygiene assessment offers a canonical instance, since hand hygiene is the single most critical defense against healthcare-associated infections (HAIs). Existing automated solutions often provide only coarse-grained compliance signals, while fine-grained technique analysis can incur substantial computational cost. Direct application of Multimodal Large Language Models (MLLMs) introduces additional challenges because processing long-form clinical videos is expensive and generic MLLMs lack grounded procedural references. We propose HH-Track, a knowledge-grounded retrieval-augmented generation (RAG) framework for fine-grained hand hygiene assessment. A skeleton-guided encoder first extracts structured clinical metrics and produces compact visual digests via boundary-progressive sampling. A prototype-anchored dual-stream RAG mechanism then retrieves step-aligned textual WHO guidelines and visual prototypes from curated reference resources. The MLLM uses these references for structured frame-level comparison during compliance assessment. To support this work, we construct Clinical-HHA, a hand hygiene dataset of 270 full-session recordings segmented into 1,408 step-level clips with template-guided expert compliance reports, paired with a curated visual prototype library. Release of both resources is subject to ongoing institutional review. Across five MLLM backbones, HH-Track improves compliance accuracy by an average of 36.1 percentage points over the Direct configurations, and the full dual-stream configuration further outperforms text-only grounding. Generated reports also show stronger agreement with expert-written references on average. These results support prototype-anchored multimodal retrieval for knowledge-grounded procedural assessment.

Graph Neural Networks (GNNs) have become a dominant paradigm for information diffusion prediction, yet their effectiveness can be undermined by biases inherent in observational diffusion data. In particular, skewed user activity and graph connectivity may encourage GNNs to exploit popularity-driven shortcuts rather than diffusion-relevant influence patterns. We attribute this limitation to the entanglement of predictive and bias-related signals in both graph structure and learned representations. To address this problem, we propose DeDiff, a disentangled graph debiasing framework that separates diffusion-relevant and bias-related signals at both the structural and representation levels. DeDiff first employs a low-rank adaptive edge mask to decompose the social and interaction graphs into complementary causal and bias views, preserving potentially informative diffusion pathways while isolating popularity-dominated shortcuts. It then introduces a dual-level representation disentanglement mechanism: causal–bias disagreement regularization discourages information leakage between the two branches, while inter-view contrastive learning preserves the distinct semantics of social and interaction relations and prevents excessive removal of useful structural information. Moreover, asymmetric branch supervision explicitly guides the bias branch to capture the empirical popularity distribution, reducing the incentive for the causal branch to encode the same shortcut and allowing it to focus on next-user prediction. Extensive experiments on four real-world datasets demonstrate that DeDiff achieves consistently strong ranking performance and remains substantially more robust than competitive baselines under amplified popularity bias.

Large language models (LLMs) have become a promising tool for scaling social experiments with silicon samples. However, fine-tuning LLMs for real randomized controlled trials faces a dual low-resource challenge: structurally balanced experimental data are often small, and the process knowledge that explains how respondent profiles and treatments jointly shape outcomes is unobserved. Without reliable process-level supervision, existing fine-tuning methods may struggle to learn why a respondent gives a particular outcome, as the model prior often provides too few useful trajectories for such domain-specific prediction tasks. To address this challenge, we propose \textbf{VPKD}, a \textbf{V}ariationally motivated \textbf{P}rocess \textbf{K}nowledge \textbf{D}istillation framework. VPKD uses a teacher LLM to generate answer-conditioned posterior trajectories, filters them through masked reconstruction to retain student-useful process knowledge, and then distills the filtered trajectories into the student via SFT. By applying VPKD to a climate-policy attitude prediction task with only 500 data points and complex profile-treatment structure, we show that the learned process knowledge not only improves prediction accuracy, but also better recovers the causal effects of the original experiment.

Detecting echo chambers is critical for understanding and limiting negative social phenomena, such as online polarization, misinformation, and conspiracy theory diffusion. However, the echo chamber detection (ECD) problem yet lacks a rigorous algorithmic foundation. We address this gap by formalizing a unified definition of echo chambers based on three principles: internal opinion homogeneity, opinion extremism, and structural isolation. Under such a definition, we establish the theoretical hardness of the ECD problem, proving it is NP-hard via a reduction from the conductance minimization problem. To circumvent this computational barrier, we derive a theoretical lower bound on the Jaccard-based homophily (JHO) of nodes that reside within valid echo chambers. This guarantee motivates JECHO, a novel algorithm that detects echo chambers via local seed expansion rather than global enumeration. JECHO first identifies seeds that exceed the JHO threshold and then applies a score-based expansion to optimize structural isolation. Extensive experiments on real-world and synthetic networks demonstrate that our theory-guided approach detects more structurally isolated echo chambers than state-of-the-art methods while reducing runtime by orders of magnitude.

The convergence of large language models and social bots allows malicious actors to manipulate the information ecosystem by generating human-like content at scale. Existing models for detecting AI-generated content often fail in the wild, primarily due to the lack of ground-truth data. We address this gap through an adversarial methodology that models the impersonation of real social media users by malicious actors. Using this methodology, we curate a multilingual, cross-platform dataset of paired human and AI-generated messages. Training on such adversarial data yields accurate detection of AI-generated text. Our approach significantly outperforms existing models for content-based bot detection in real-world, out-of-distribution data.

Social bot detection is essential for protecting online platforms from misinformation amplification, coordinated manipulation, and distorted public discourse. However, large language models have made social bots much harder to detect from text alone because semantic camouflage is now cheap, fluent, and scalable. The resulting challenge is modality conflict: an account may look human-like in semantics while remaining suspicious in graph structure, profile attributes, or cross-modal consistency. Recent graph-based detectors tackle this limitation by adding graph-side complexity, such as sparse prototype selection, adaptive gating, or architecture-specific control logic, yet our experiments suggest that complexity alone is not the most reliable way to resolve such conflict. We therefore propose CSC, a calibrated-simplicity framework for conflict-aware LLM-era social bot detection. The framework combines three design choices: a simplified prototype-guided graph expert that retains useful structural biases while removing unstable graph-side heuristics, calibrated simplex-constrained fusion that aligns heterogeneous confidence spaces before late fusion, and a lightweight inconsistency expert that models cross-modal disagreement. Experiments on TwiBot-22, TwiBot-20, and MGStBot-large show that CSC improves calibrated operating-point decision quality while remaining competitive across external benchmarks. Further analyses show that calibration improves confidence reliability, the inconsistency expert mainly provides localized corrections in high-conflict or near-threshold regions, and simplified graph-side control yields a better stability-cost trade-off. A targeted semantic-camouflage stress test further shows that replacing selected bot text with matched human text sharply degrades the standalone text expert while leaving graph and fused evidence stable on a balanced challenge set. Overall, these results suggest that, when semantic evidence is easy to camouflage, robust bot detection benefits from preserving reliable structural signals and calibrating conflicting modalities before fusion, rather than continually increasing graph-side complexity.

To detect social bots, many advanced methods have been proposed, especially graph-based approaches, their effectiveness is undermined by sophisticated bots that create spurious associations in network structures. As a result, current methods often struggle to provide high-confidence detection results. This motivates us to quantify the uncertainty of the output and reduce the interference of misinformation on detection, so as to determine the confidence of the results and output high-confidence results. Therefore, we propose an uncertainty-aware bot detection method to inform the confidence and use the uncertainty score to pick a high-confidence decision from multiple views of a social network under different environments. Specifically, our proposed BotUmc uses LLM to extract information from tweets. Then, we construct a graph based on the extracted information, the original user information, and the user relationship, and generate multiple views of the graph by intervention. Lastly, an uncertainty loss is used to force the model to quantify the uncertainty of results and select the result with low uncertainty in one view as the final decision. Extensive experiments show the superiority of our method.

Illicit account detection in blockchain-based cryptocurrencies seeks to pinpoint anomalies within large-scale, decentralized transaction networks, figuring prominently in financial risk control systems. Existing efforts typically formulate transaction networks as directed multigraphs and leverage graph neural networks (GNNs) to capture topological information under limited supervision. Despite their decent performance, supervised GNN models are capped by data scarcity and structural noise, with the former arising from the lack of annotated data and the latter stemming from malicious actors forging deceptive connections. In light of these, we in this paper propose LABRA, a novel \underline{la}bel-free graph modeling for detecting illicit accounts in \underline{b}lockchain t\underline{ra}nsaction networks. Operating in an unsupervised manner, LABRA\footnote{The source code and datasets are publicly available at https://github.com/bicheng1225/LABRA.} leverages a discrepancy-aware graph encoder to produce discriminative node representations, and subsequently employs an anchor-guided scoring mechanism to designate anchor nodes for anomaly detection. Furthermore, a tailor-made message-passing scheme is introduced to mitigate structural noise, which captures atypical financial flows within a transaction graph while attenuating the feature-smoothing effect caused by high-degree nodes. Extensive experiments on several cryptocurrency benchmark datasets demonstrate the feasibility and superiority of our method over several cutting-edge methods. To our knowledge, LABRA is among the first training-free frameworks for illicit account detection in cryptocurrencies, delivering robust performance without resorting to any learnable parameters or iterative model optimization.

Predicting stock returns requires modeling persistent company relations and time-varying market signals. Graph Neural Network (GNN) approaches combine static priors with dynamic relational graphs, but treat every extracted edge as a trustworthy binary fact, ignoring heterogeneous reliability across relational sources. We propose \textbf{LENS} (\textbf{L}LM-\textbf{E}xtracted priors and \textbf{N}ews-driven \textbf{S}emantics), a static--dynamic graph framework for stock return prediction and ranking that models relational reliability as a first-class signal across the pipeline. On the static side, we extract a multi-relational company graph from a Large Language Model and aggregate it over multiple runs with confidence weighting to downweight low-consistency edges. On the dynamic side, we build a news-driven semantic graph with exponential-decay temporal aggregation to capture changing market context under uneven coverage. A context-dependent gated fusion module rebalances the two graphs at the per-stock, per-day level. On S\&P 500, LENS improves AR@5 by 16.4\% and SR@5 by 22.6\% over the strongest baseline, and on NASDAQ 100 it achieves comparable improvements (15.5\% and 20.9\% respectively). Ablations isolate each component's contribution, and a human-validation study of sampled relations supports that the aggregation is grounded, providing evidence that reliability-aware LLM relations and news-based semantic graphs carry complementary information for graph-based financial prediction.

Federated continual learning (FCL) suffers from complex spatio-temporal catastrophic forgetting, driven by distribution drift along task streams and aggregation conflicts under non-IID client data. Existing studies often treat temporal and spatial forgetting as decoupled objectives, overlooking their interaction in the overall memory dynamics of the federation, where client-side acquisition and server-side consolidation jointly shape long-term performance. Motivated by the structural and functional correspondence between the dual-channel memory mechanism in collective cognition and FCL's client-server collaboration, we introduce Federated System Memory (FSM), a system memory metric that jointly models client-side short-term retention and server-side long-term consistency. Based on FSM, we further propose FedMem, a memory-aware FCL framework that utilizes dual-channel memory control to enable system-level closed-loop optimization. On the client side, FedMem constrains local learning with dual-channel decay indicators: short-term decay strengthens prototype-based retention, while long-term decay enhances memory-guided regularization for stable task-memory retrieval. On the server side, it regulates global updates through collective-consistency gated aggregation and adaptive update scaling based on long-term decay, while maintaining global prototypes and memory. Extensive experiments demonstrate that FedMem consistently outperforms existing methods in both learning performance and system memory retention.

Prototype-based federated learning (FL) provides a promising route for federated heterogeneity. By exchanging class prototypes rather than model parameters, it provides class-level semantic guidance under non-IID data and accommodates heterogeneous client backbones. Under severely skewed label distributions, however, two coupled limitations emerge. The broadcast global prototype is a cross-client average, and direct adoption can cause client semantic dilution. Meanwhile, local prototypes of rare classes are estimated from a few samples, leading to local margin erosion against well-supported neighbours. To address both, we propose ProRC, a prototype personalization framework grounded in shrinkage estimation. ProRC replaces naive alignment with a reliability-calibrated, margin-aware interpolation between global and local prototypes, and comprises two complementary client-side modules. The first module, Reliability-Calibrated Prototype Adaptation (RaPA), calibrates each global prototype by the reliability of its local counterpart, restoring client-specific semantics only to the extent the data warrant. The second module, Margin-Aware Prototype Refinement (MaPR), audits the candidate against the local true-class margin and attenuates margin-reducing steps. Together, the two modules are designed to mitigate client semantic dilution and local margin erosion, while preserving the prototype-exchange protocol and per-round payload of prior methods. Extensive experiments on label-skewed and real-world image benchmarks, together with ablation and reliability analyses, show that ProRC achieves the best overall average accuracy among the compared methods and improves the strongest prototype baseline in most heterogeneous settings.

Collaboration-based personalized federated learning (Co-PFL) mitigates data heterogeneity by enabling clients to selectively exchange knowledge through a learned collaboration graph. Despite their success, existing Co-PFL methods often update collaboration weights mainly from pairwise similarity, which can produce graphs that are either overly sparse and hinder knowledge propagation or overly dense and weaken personalization, leaving graph connectivity largely uncontrolled. In this paper, we propose FedPGP, a Co-PFL framework that explicitly schedules collaboration from dense to sparse over training rounds, enabling clients to first exploit broadly shared knowledge and then refine specialized knowledge aligned with their local distributions. FedPGP builds on a connectivity-aware similarity regularization that jointly accounts for global graph connectivity and client-model similarity, together with an efficient greedy pruning rule that removes edges with the least contribution to the regularized objective. We further present FedPGP+, a module-specific extension that learns different collaboration graphs for different model modules, allowing collaboration intensity to vary along model depth. Extensive experiments across diverse non-IID settings show that FedPGP and FedPGP+ consistently outperform state-of-the-art Co-PFL baselines, achieving up to 5.52% absolute accuracy improvement, while producing interpretable dense-to-sparse collaboration schedules.

Federated Learning (FL) provides a privacy-preserving paradigm for collaborative learning across distributed clients. In practical FL systems, clients often exhibit non-IID data distributions and deploy diverse model architectures under resource constraints. While model heterogeneity improves deployment flexibility under client resource constraints, it prevents direct parameter aggregation and makes cross-architecture knowledge transfer challenging. Existing methods mainly rely on shared structures, coarse-grained distillation, or projection-based feature alignment, which cannot explicitly capture channel-level semantic correspondences across heterogeneous architectures. This paper proposes FedCSOT, a model-heterogeneous personalized FL framework based on channel semantic optimal transport. FedCSOT induces sequence-driven channel semantics from channel responses over a shared unlabeled sequence, establishing a common semantic space for cross-architecture alignment. The framework models cross-architecture semantic alignment through entropy-regularized optimal transport to facilitate knowledge transfer. It further designs a divergence-aware personalized fusion strategy to adaptively balance consensus knowledge and local personalization under non-IID data distributions. Extensive experiments under diverse data and model heterogeneity settings demonstrate that FedCSOT consistently improves average test accuracy, convergence behavior, and robustness over representative baselines. The code is available at \url{https://github.com/greenhead01/FedCSOT}.

Customizing Pretrained Models (PMs) for downstream tasks demands task-specific data, which burgeoning Web3 applications provide abundantly. Blockchain-powered federated learning (BCFL) offers a secure and privacy-preserving paradigm to customize PMs among \textit{self-interested} and \textit{untrusted} Web3 participants. Yet, existing BCFL algorithms face severe \textit{efficiency} and \textit{scalability} challenges when applied to large PMs, due to their substantial computational and memory requirements and the \textit{hardware constraints} of many participants. This paper introduces \texttt{DICE}, a novel BCFL algorithm that addresses the efficiency and scalability challenges of customizing large PMs in Web3 via two sets of synergistic techniques: (i) a \textit{dynamic progressive learning} scheme and a \textit{model slice verification} scheme to enable a wider range of honest participants, including those unable to deploy the full model, to engage in customizing PMs, while preserving model utility; (ii) a \textit{tree-structured committee consensus} protocol and an \textit{efficiency-driven incentive} scheme to address the dominant bottleneck that hinders existing BCFL algorithms from scaling to a larger number of data holders for customizing data-hungry large PMs, while maintaining security. Theoretical analysis and extensive experimental results on customizing vision and language PMs demonstrate \texttt{DICE}'s efficiency, effectiveness, and security. A comprehensive technical report of \texttt{DICE} including its implementation details, extended analyses, discussions, and experiments is available at \url{https://github.com/Hoho-Lin/DICE}.

Federated Learning is a promising paradigm to achieve privacy-preserving collaborative training. However, its performance degrades severely under data heterogeneity. A main obstacle is the divergent local representations induced by data heterogeneity, which will degrade the quality of the aggregated global representations. In this paper, we discover that although local representation spaces contain client-specific biases, their internal inter-class structure remains consistent across clients, which we identify as relative consistency. Through theoretical and empirical analyses, we further show that a proxy dataset can serve as a common anchor to translate class prototypes into a relative space, where client-specific shifts are suppressed and cross-client class centers become better aligned. Based on these findings, we propose a novel framework named \textit{Federated learning with Relative Representation Alignment} (FedReRA) with Relative Knowledge Extraction (RKE) and Relative Ensemble Inference (REI) to extract and leverage the relative consistency to address data heterogeneity. Specifically, RKE debiases the prototypes by translating absolute prototypes into the aligned relative space via a flexible proxy dataset, while REI conducts collaborative inference with bias-reduced prediction in the relative space. Extensive experiments validate FedReRA’s superiority over state-of-the-art methods, achieving a $3.83\%$ average improvement under high heterogeneity and notably sustaining nearly $10\%$ improvement on CIFAR-100 even when using random noise as the proxy data.

Federated learning under non-IID data suffers from client drift. Recent work introduces Mixture-of-Experts (MoE) architectures to handle data heterogeneity by routing inputs to specialized experts, but two problems remain. Softmax routing produces overconfident expert assignments with no reliability measure. The parameter space carries no class-level information, forcing the server to aggregate all clients uniformly and destroying the expert specialization built during local training. We call this class-level specialization invisibility. To address these problems, we propose FedEMoE. On the client side, Evidential Mixture-of-Experts equips router and experts with evidential heads, dynamically adjusts the number of active experts based on routing uncertainty, and compresses accumulated evidence into a per-client signature. On the server side, Evidence-Guided Class-Aware Aggregation takes the signatures as input and adaptively performs class-aware parameter fusion and component recombination. Experiments on CIFAR-10, CIFAR-100, and FEMNIST show that FedEMoE consistently outperforms eight baseline methods across all settings. Our code is available at https://github.com/Stephen-Chow1/FedEMoE-CIKM.

Robust aggregation in federated learning is often framed as down-weighting abnormal client updates. However, in realistic cross-device regimes, each round observes only a small number of high-dimensional updates, while benign updates can already be anisotropic, heavy-tailed, and strongly affected by non-IID data. In this setting, robustness depends not only on the final weighting rule, but also on whether the server-visible update geometry is estimable. We propose FedBIG (\textbf{Fed}erated \textbf{B}ounded-\textbf{I}nfluence \textbf{G}eometry), a server-side aggregation framework that separates geometry construction from influence control. Instead of scoring raw updates directly, FedBIG builds a stabilized sketch-space geometry using robust centering, low-dimensional projection, calibrated scatter estimation, shrinkage, and whitening. It then maps stabilized deviations to bounded-influence weights and aggregates the original client updates. Experiments on heterogeneous federated learning benchmarks show that FedBIG improves the robustness--utility trade-off under statistical and triggered poisoning. Ablations, boundary stress tests, and runtime analysis further support the central claim: robust aggregation is effective when server-visible geometry is estimable, and weakens when this geometry collapses under extreme heterogeneity or sparse participation. Our code is available at https://github.com/ytma-phoenix/FedBIG.

Document databases provide flexible data modeling, but long-running applications still rely on implicit schemas that evolve continuously. Legacy documents do not automatically follow schema changes, so a collection may contain documents conforming to different historical schema versions. This makes schema evolution an execution problem that spans change representation, operation reconstruction, and legacy-data migration. Existing studies typically address these tasks separately, leaving schema-level changes disconnected from the instance-level migration decisions. This paper proposes an operation-guided framework for executable schema evolution and migration in document databases. We first define a unified operation framework over field paths, types, and constraints, with explicit User-Defined Function (UDF) contracts for destructive changes. This framework specifies the valid operations from which schema evolution sequences are composed. We then reconstruct executable schema evolution sequences from schema snapshots through a two-stage method that combines multi-channel field-evolution association with instruction-tuned sequence generation. Finally, the generated sequences drive progressive migration: an instance-level Benefit--Cost Optimization (BCO) selector prioritizes affected documents, while a window-level allocator distributes migration budget under live workloads. Extensive experiments demonstrate that the proposed framework directly expresses all 32 curated evolution cases, improves end-to-end sequence executability by 29.44\%--31.51\%, and reduces residual migration debt by 24.73\% over the strongest adaptive migration baseline.

\begin{abstract} Filtering textual columns based on user intent is a pervasive task in database applications, yet discrepancies between user predicates and concrete database values often cause keyword searches to fail. The Recall-Then-Verify (RTV) paradigm, which first retrieves a budget of candidate values and then verifies them using Large Language Models (LLMs), has emerged as a promising and widely adopted strategy; however, its overall efficacy remains inherently bounded by the recall of the initial retrieval stage. To improve this recall, we propose \methodname, a type-specific query reformulation framework that adaptively aligns predicates with database content by addressing database semantic gaps. Specifically, \methodname categorizes these semantic discrepancies into three distinct types—synonymy, conceptual hierarchies, and format inconsistencies—and executes a specialized, type-specific reformulation strategy tailored to each gap. \methodname further grounds reformulation in the target database by reusing LLM-verified column values as feedback, enabling the system to capture database-local variants. Furthermore, to maintain low latency, \methodname introduces an on-demand triggering mechanism that activates LLM reformulation only when a significant semantic gap is detected, supplemented by adaptive selection and early stopping strategies to prune redundant computation. Compared to existing RTV methods, experimental results demonstrate that \methodname consistently enhances retrieval recall by up to 45.45\% and improves end-to-end F1 score by up to 26.72\%. \end{abstract}

The existence of inconsistencies in databases has motivated the study of Consistent Query Answering (CQA) since the 1990s. Recent work proposed the use of Maximum Satisfiability (MaxSAT) for solving CQA for concrete types of inconsistencies. This paper shows that, for the types of inconsistencies solved with MaxSAT, it is the case that CQA corresponds to finding the backbone of a propositional formula. This observation not only yields a novel characterization in terms of the computational complexity of the problem, but it also enables significantly more efficient practical solutions. The experiments confirm that significant performance improvements can be obtained by computing the backbone instead of solving MaxSAT.

Time-window aggregate queries over temporal graph streams—COUNT, COUNT_DISTINCT, and label-filtered predicates—are central to agent memory, communication monitoring, and knowledge-base analytics, yet exact evaluation becomes increasingly costly as timestamped edges accumulate. Sketch-based approximate query processing (AQP) offers compact summaries, but the standard per-bucket design couples temporal resolution with cardinality state: placing a HyperLogLog (HLL) inside every Count-Min Sketch (CMS) bucket leaves only coarse time partitions under a fixed budget and still does not natively support all three query types. We present the Cardinality-Aware Multi-Resolution Sketch (CAMRS), a budget-decoupled multi-resolution sketch built on the observation that CMS and HLL have distinct budget elasticities. CAMRS separates fine-grained CMS buckets from coarse HLL segments, achieving a 9.6× temporal-resolution gain under the same 1 MB budget, and augments each bucket with Space-Saving label tracking to support predicate queries without global label thresholds. We further derive a closed-form COUNT error bound showing that temporal resolution, rather than CMS width, is the dominant accuracy lever; this bound drives a reliability-aware fallback mechanism that controls tail risk when predicted sketch error exceeds a user tolerance. Experiments on six real-world temporal graphs (60K–17.8M edges) show that CAMRS reduces COUNT error by up to 29.7× and predicate error by up to 7.9× over the coupled baseline while preserving native COUNT_DISTINCT support. The adaptive fallback policy further reduces worst-case tail error by up to 667×.

Ontology-based data access is intended for graph data, but practical support remains limited to SQL-like languages lacking the navigational and path-matching features fundamental to graph querying. Theoretical algorithms for navigational queries have long been available. Still, they have never been implemented, largely because practical graph query languages fell short of their theoretical counterparts and lacked the expressive power needed to support rewriting common ontology languages. The recent standardisation efforts around GQL and SQL/PGQ finally allow us to overcome this barrier and present a practical query rewriting technique for ontology-mediated navigational graph queries. Ontologies are written in a DL-Lite variant tuned to property graphs, and the queries in a GQL fragment with nested two-way regular path queries and which can be evaluated in Cypher. Preliminary experiments with our proof-of-concept prototype suggest that querying graph data with ontological knowledge may finally be within reach.

As property graphs become central to domains like network analysis, legislative and financial monitoring, there is an urgent need for reactive processing capabilities that can respond to complex structural changes. At the same time, the standardization of property graph processing has culminated in the ISO Graph Query Language (GQL) standard, which establishes a uniform and portable foundation for expressing triggers, ensuring portability across implementations. This paper introduces GQL-Triggers, the first trigger framework fully aligned with the ISO GQL standard that treats paths as first-class citizens. Unlike traditional approaches, GQL-Triggers allow developers to define declarative events, conditions, and actions over complex path patterns, incorporating GQL-specific features. We provide a formal semantic model for path-level event detection and transition sets computation. We present two alternative implementations of GQL-Triggers on top of Neo4j, respectively named Index-based and Query-based, and provide their experimental comparison in terms of time and memory efficiency.

Continuous subgraph matching over dynamic graphs is a fundamental problem in graph databases, with broad applications in real-time pattern monitoring and risk control. In practice, applications often require monitoring multiple graph patterns simultaneously, making multi-query oriented continuous subgraph matching (MCSM) an increasingly important research topic. Although existing methods have achieved performance gains, their indexes offer limited filtering power and they lack high-quality query plans tailored to the multi-query setting. To address these limitations, we propose Cimos, a novel index-based framework for cost-effective MCSM. First, we design a data-centric index called DCL that serves multiple query graphs and prunes more non-candidate vertices by jointly considering the matching constraints of ancestor and descendant vertices in query DAGs. Second, we adopt a hybrid-order planning strategy that separates the query plan into static and dynamic ordering components. For the static component, we introduce a prefix order tree to extract isomorphic subgraphs shared across query graphs, and optimize vertex orderings using statistical upper-bound obtained via DCL-based subgraph sampling. For the dynamic component, we employ DCL-based cardinality estimation to adaptively determine vertex orderings during the matching process. Extensive experiments demonstrate that Cimos achieves notable speedup of more than 4.2x against state-of-the-art CSM and MCSM methods.

Learned indexes have significantly improved the query performance of in-memory databases by transforming traditional structural traversals into cumulative distribution function model predictions. However, existing dynamic learned indexes rely heavily on the core assumption of "local linear data distribution." Under highly skewed dynamic workloads in real-world scenarios, this linear assumption frequently fails, forcing the system to trigger extremely expensive model retraining and horizontal data shifting. This incurs severe cascading splits and leads to a sharp degradation in dynamic write throughput. To break this architectural bottleneck, this paper proposes BLADE, a bit-level learned index architecture designed for dynamic mixed workloads. BLADE abandons the strong reliance on complex mathematical models and achieves a structural decoupling of routing addressing and underlying data storage at the physical level. In the upper routing module, BLADE employs a pure integer bitwise partitioning mechanism for stateless memory addressing, eliminating the overhead of floating-point model computation. In the underlying leaf nodes, the system inherits the contiguous array layout with reserved gaps to absorb insertions efficiently. To address local congestion caused by data skew, BLADE introduces a vertical splitting strategy, which redistributes data by extending the routing level downwards, effectively avoiding large-scale physical data copying at the bottom layer. Comprehensive experimental evaluations demonstrate that under read-only and read-heavy workloads, BLADE achieves point lookup throughput of 8.2$\times$, 4.1$\times$, 3.2$\times$, and 1.3$\times$ that of B+-trees, PGM-index, ART, and ALEX, respectively. Under high-frequency write dynamic mixed workloads, its throughput reaches up to 4.5$\times$, 4.0$\times$, 2.3$\times$, and 1.2$\times$ that of PGM-index, B+ Tree, ART, and ALEX. Furthermore, by replacing learned internal-node models with deterministic bitwise routing, BLADE eliminates the iterative model fitting overhead, accelerating the index initialization and bulk-loading speed by 16$\times$ to 48.8$\times$ compared to the most advanced learned indexes. Under the tested skewed insertion patterns, BLADE's maximum observed RSS is only 1.17$\times$ that of its Uniform baseline. Parameter-sensitivity experiments further validate the default 30\% target gap and 16~MiB maximum leaf capacity as a balanced configuration for insertion performance and space efficiency.

Effective personalization on large-scale job platforms requires modeling members based on heterogeneous textual sources, including profiles, professional data, and search activity logs. As recommender systems increasingly adopt Large Language Models (LLMs), creating unified, interpretable, and concise representations from heterogeneous sources becomes critical, especially for latency-sensitive online environments. In this work, we propose a novel Reinforcement Learning (RL) framework to synthesize a unified textual representation for each member. Our approach leverages implicit user engagement signals (e.g., clicks, applies) as the primary reward to distill salient information. Additionally, the framework is complemented by rule-based rewards that enforce formatting and length constraints. Extensive experiments across multiple LinkedIn products serving one of the world’s largest job platforms demonstrate significant improvements in key downstream business metrics. This work provides a practical, labeling-free, and scalable solution for constructing interpretable user representations that are directly compatible with LLM-based systems.

Large Language Models (LLMs) can provide semantic reasoning for recommendation, but direct LLM invocation is incompatible with the latency, stability, and reliability requirements of industrial ranking systems. We propose \textbf{RA-CoT} (Retrieval-Augmented Chain-of-Thought), a deployment-oriented framework that uses LLM-derived structured reasoning without replacing the production CTR/CVR backbone or placing LLM inference on the blocking serving path. RA-CoT asynchronously generates schema-constrained CoT-derived representations from pre-impression behavior summaries, caches them with validity checks, and falls back to the backbone when no valid reasoning vector is available. Valid reasoning is injected through an availability-aware gated residual module as bounded logit-space compensation, making the LLM signal reusable, safely ignorable, and compatible with millisecond-level serving. We evaluate RA-CoT on a large-scale industrial advertising dataset and a public recommendation benchmark, and deploy it in a two-week online A/B test. RA-CoT yields modest but consistent offline ranking improvements across DNN, DIEN, BST, and SIM backbones. In production, it improves CTR by 2.48\%, CVR by 1.95\%, and RPM by 2.12\%, while keeping P99 serving latency at 54 ms. These results show a practical path for using cached and bounded LLM-derived structured reasoning in industrial recommender systems.

In real-world recommender systems, different retrieval objectives are typically addressed using task-specific datasets with carefully designed model architectures. We demonstrate that Large Language Models (LLMs) can function as universal retrievers, capable of handling multiple objectives within a generative retrieval framework. To address the model expressiveness and inference efficiency issues of generative retrieval, we propose multi-query representation and transferable discriminative mapping layers to model complex user-item relationships, and incorporate probabilistic sampling to reduce computation costs. Finally, our Universal Retrieval Model (URM) can adaptively generate a set from tens of millions of candidates based on an arbitrary given objective while keeping the latency within tens of milliseconds. Applied to industrial-scale data, URM outperforms expert models elaborately designed for different retrieval objectives in offline experiments and achieves a $3.13\%$ revenue increase on our online advertising platform. URM also demonstrates strong performance in impression and GMV for long-tail and cold-start items.

Slate recommendation, which presents users with a ranked item list in a single display, is ubiquitous across mainstream online platforms. While recent generative recommendation methods have shown strong potential in modeling item sequences with semantic IDs, directly applying them to industrial-scale slate recommendation faces a fundamental disconnect: entangled SID spaces confound high-level list planning, fine-grained autoregressive decoding over long sequences limits semantic planning efficiency, and token-level objectives misalign with holistic slate quality. In this paper, we propose HiGR, an industrial-scale hierarchical generative framework for slate recommendation that bridges this disconnect through a co-designed pipeline. First, HiGR learns structured SIDs via a Prefix-Contrastive Residual Quantized VAE (PCRQ-VAE). By enforcing high-level prefixes to capture shared semantics, PCRQ-VAE creates a controllable discrete space that acts as a prerequisite for efficient planning. Leveraging this structured space, our Hierarchical Slate Decoder (HSD) shifts autoregressive modeling from entangled token-level decoding to coarse-grained preference embeddings. This design significantly reduces inference latency while allowing explicit global slate structure planning. Finally, this stable planning space enables an ORPO-based listwise alignment mechanism to optimize triple-objective implicit feedback—ranking fidelity, genuine user interest, and diversity. Extensive offline experiments show that HiGR outperforms state-of-the-art baselines by over 10\% in offline recommendation quality while achieving a $5\times$ inference speedup. Online A/B tests on Tencent platforms further improve average watch time by 1.22\% and average video views by 1.73\%. HiGR has been deployed on multiple Tencent platform surfaces (e.g., news, videos, and web novels), serving hundreds of millions of users and proving its industrial-scale applicability.

Positional encoding is a fundamental component of Transformer-based generative recommendation models, where user histories are modeled as autoregressive item sequences. Most positional encoding methods are inherited from natural language processing and mainly represent discrete item order. However, recommendation sequences go beyond ordered lists, as timestamps and temporal effects also shape item relations. Our work is motivated by a real-world food delivery and instant retail recommendation system, where user behavior exhibits multi-level temporal regularities, including recency effects, meal-time peaks, weekday-weekend shifts, and promotion-driven traffic bursts. Existing methods partially address this issue through timestamp features, interval embeddings, decay functions, or attention biases, but they usually inject heterogeneous temporal signals through a unified representation or a single modeling pathway, making it difficult to distinguish broad temporal dynamics from local order cues. To address this limitation, we propose Decoupled Temporal Encoding (DTE), a lightweight framework for generative recommendation. DTE separates temporal dynamics from order information through two complementary modules: a personalized macro-temporal module that injects compact temporal primitives into item embeddings, and a time-gated micro-sequential module that introduces relative-order bias only when interactions are temporally dense. DTE is also parameter-efficient and deployment-friendly, allowing easy integration into existing systems. Offline experiments on public and industrial datasets demonstrate DTE’s superior performance. Online A/B tests further show improvements of +1.8% Click-through Rate (CTR) and +3.0% Revenue Per Mille (RPM) with only +0.3% serving latency overhead over the baseline positional encoding scheme. DTE has been deployed to the full Shangou ad recommendation traffic in the Taobao App since January 2026, highlighting the practical value of decoupling macro-temporal context from sequential order in real-world generative recommendation.

Generative recommendation has emerged as a transformative approach in industrial-scale systems, yet faces critical challenges in production environments, including Semantic Drift/Gap and cold start, and preference inconsistency. As NetEase's game platform serving more than 6 Million users, we identified that hash ID-based generative models suffer 84.6\% performance degradation on the 60\% long-tail items that generate minimal revenue but are crucial for user retention. Although multimodal content features could enrich item representations via semantic IDs, extracting and storing them can be tens of times higher than compact semantic IDs. Moreover, existing semantic IDs can be inconsistent with collaborative signals in industrial scenarios. To address this limitation, we propose LASER, a Label-Anchored Semantic Encoding with Multimodal-Collaborative Signals Harmonization for Industrial-Scale Generative Recommendation. LASER introduces two novel components: (1) Multimodal Semantic ID Learning, which constructs semantic IDs from multimodal features and aligns them with collaborative signals, via label-driven residual quantization and Swing-weighted pairwise alignment; and (2) Preference-aligned generative sequential recommendation, which leverages hierarchical semantic IDs to improve robustness under sparse interactions and cold-start, via popularity-aware fusion and preference-aligned autoregressive training. We conduct extensive experiments on real-world datasets and industrial scenarios to demonstrate the effectiveness of our model.

Item tokenizer encodes semantic embeddings into token IDs to replace the randomly assigned item IDs used in traditional recommendation models, fundamentally addressing the problems of excessive parameters and cold starts. However, the most common tokenizer, RQ-VAE, suffers from low decoding efficiency due to the inherent dependencies among its codebooks. Meanwhile, efficient independent tokenizers such as optimized product quantization (OPQ) still struggle with dimensional correlations and distribution complexity of semantic embeddings. In this work, we propose a f\underline{low}-based item \underline{T}okenizer (Tlow) to transform raw semantic embeddings into a latent space where embeddings conform to a unified standard normal distribution, achieving dual advantages of dimensional independence and distributional simplicity. Independent tokenization performed on these latent embeddings yields semantically clear token IDs. Additionally, we introduce a novel codebook guidance to align the codebook space with the token embedding space, further aiding the learning of more semantically distinct token embeddings. Offline experiments on four public datasets demonstrate that Tlow's tokenization and codebook guidance significantly improve recommendation performance. The improvement on cross-domain and multi-modal recommendations also proves the effectiveness of item tokenization in a simplified embedding space. Online experiments for a multi-modal retrieval task on China's largest social media platform WeChat validate Tlow's powerful distribution transformation capability. The retrieval model based on token IDs improves user CTR by 10.32\% globally and by 11.64\% for new items. Our codes are available at \url{https://github.com/wjjln/Tlow}.

Traditional ID-based recommender systems often struggle with cold-start and generalization challenges. Multimodal recommendation systems, which leverage textual and visual data, offer a promising solution to mitigate these issues. However, existing industrial approaches typically adopt a two-stage training paradigm: first pretraining a multimodal model, then applying its frozen representations to train the recommendation model. This decoupled framework suffers from misalignment between multimodal learning and recommendation objectives, as well as an inability to adapt dynamically to new data. To address these limitations, we propose LEMUR, the first large-scale multimodal recommender system trained end-to-end from raw data. By jointly optimizing both the multimodal and recommendation components, LEMUR ensures tighter alignment with downstream objectives while enabling real-time parameter updates. Constructing multimodal sequential representations from user history often entails prohibitively high computational costs. To alleviate this bottleneck, we propose a novel memory bank mechanism that incrementally accumulates historical multimodal representations throughout the training process. In a 14-day online A/B test on Douyin Search, LEMUR has led to a significant 0.843% reduction in query change rate decay, a 0.558% increase in staytime and a 0.81% improvement in QAUC. Additionally, LEMUR has shown significant gains across key online or offline metrics for Douyin Advertisement, Tiktok Search and Ecom Search. Our results validate the superiority of end-to-end multimodal recommendation in real-world industrial scenarios.

Multimodal representations have been widely adopted in industrial e-commerce recommendation systems. Due to their strong semantic understanding and generalization capabilities, they enhance the performance of traditional sparse ID-based Click-Through Rate (CTR) prediction models. Current multimodal application frameworks in the CTR prediction task typically follow a two-stage paradigm: first, pre-training a multimodal encoder on data from specific recommendation scenarios; second, extracting items' multimodal representations using this pre-trained multimodal encoder and integrating them into the CTR prediction model. However, the training objectives and data distribution of multimodal pre-training tasks often differ from those of the CTR prediction task, which limits the effectiveness of multimodal representation on downstream tasks. In this paper, we focus on how to learn Native Multimodal Representation for the CTR prediction task. One intuitive solution is to jointly train the multimodal encoder and CTR model end-to-end on the CTR task, with the expectation that the encoder can automatically learn downstream-relevant knowledge. However, we find that the end-to-end training does not bring performance improvements to existing multimodal application paradigms. Our analysis reveals that user behaviors in raw CTR data are driven by both multimodal semantics and non-multimodal factors, leading to ambiguous supervision and inconsistent encoder updates. To address this, we propose a Mine-Then-Train method that mines high-quality, multimodally interpretable training samples from CTR data and uses them to fine-tune the multimodal encoder for better alignment with user click preferences. Offline and online experiments demonstrate the effectiveness of our approach.

\begin{abstract} The zero-inflated and extreme long-tailed nature of Lifetime Value (LTV) data poses a dilemma for existing methods: ranking-oriented methods lack fine-grained value estimation, while regression-oriented methods struggle to converge in high-variance spaces. We propose BasisLTV, a novel framework that integrates efficient semantic representation with robust value reconstruction. To address feature sparsity and computational bottlenecks, we introduce Efficient Semantic Behavior Compression. This mechanism compresses high-dimensional attribute tuples into compact, attribute-aware behavior representations, enriching input context while maintaining low inference costs. Building on this, BasisLTV tackles long-tailed regression via Ordinal-Guided Basis Reconstruction. Specifically, it imposes macro-constraints through ordinal classification and reconstructs micro-values using interval basis vectors. Furthermore, a Null-Space Constraint is introduced, leveraging a null-space vector strictly orthogonal to the reconstruction basis to decouple out-of-interval noise. Experiments on both a real-world industrial dataset of 10 million users and a public benchmark demonstrate that BasisLTV significantly outperforms state-of-the-art methods in both ranking and regression metrics. Furthermore, online A/B testing verifies its robust efficacy and significant practical value in a real-world advertising system. \end{abstract}

Accurately predicting conversion rates (CVR) for low-activity users remains a fundamental challenge in large-scale e-commerce recommender systems. For these users, existing approaches face three critical limitations: (i) data dominated by high-activity users may introduce noisy and mismatched signals for learning low-activity users' preferences; (ii) insufficient interaction information; and (iii) a systemic training bias toward high-activity users. To address these challenges, we propose \textit{ChoirRec}, a novel framework that uses Large Language Models (LLMs) to identify semantically similar users and transfer reliable knowledge from richer user groups to low-activity users.ChoirRec comprises three components: (i) a Semantic Group Generation module that utilizes LLMs to form reliable, cross-activity user clusters, thereby filtering out noisy signals; (ii) a Hierarchical Representation module that enriches sparse user embeddings with informative group-level priors to mitigate data insufficiency; and (iii) a Multi-granularity Module that employs a dual-channel architecture and adaptive fusion mechanism to ensure effective learning and utilization of group knowledge. We conduct extensive offline and online experiments on Taobao, a leading industrial-scale e-commerce platform. ChoirRec improves GAUC by 1.16\% in offline evaluations, while online A/B testing reveals a 7.24\% increase in order volume, highlighting its substantial practical value in real-world applications.

On billion-scale live streaming platforms such as TikTok, a core challenge is to distribute millions of newly initiated live streams to potential viewers. The ephemerality and real-time variability of live streams demand an aggressive and dynamic cold-start strategy to capture value within an extremely limited window. Beyond this temporal urgency, the system faces a unique challenge: a two-way interactive feedback loop in which recommendations directly influence streamer behavior and subsequent viewer feedback. Traditional methods, however, lack the ability to intervene at critical learning moments and fail to provide the active incentives needed to optimize the streamer-user feedback loop. In this paper, we propose LIFT (Learning-aware Intervention Framework for Cold-Start Recommendation on TikTok Live), a novel paradigm designed specifically for this real-time intervention. LIFT employs three adaptive modules to optimize traffic allocation throughout the cold-start phase: (1) Phase-Aware Exploration Module, which identifies optimal learning windows by analyzing the dynamics of item embeddings, enabling precisely timed exploration of live streams; (2) Incentive-Driven Interaction Module, which distills LLM-derived interaction quality signals into the recommender system to prioritize targeted streams, effectively incentivizing content-rich interaction; (3) State-Based Gating, which determines whether the current state of a live stream merits exposure within the dynamic listwise context, ensuring that interventions are both reasonable and efficient. Extensive offline and online experiments demonstrate that LIFT outperforms standard production baselines with statistically significant improvements. Since its full deployment on TikTok, the framework has achieved an evaluated 8.91 increase in live viewing duration and an observed 9.65 increase in accumulated stream duration. These results validate that our proactive, real-time intervention strategy yields effective performance, given the unique nature of live streams.

Federated recommender systems (FedRec) aim to preserve user privacy by keeping sensitive data on client devices and sharing only model parameters with a central server. However, FedRec is still vulnerable to attribute inference attacks (AIAs), where server-side adversaries exploit uploaded parameters to infer users’ private attributes. Existing FedRec approaches face a suboptimal privacy-performance trade-off. Privacy-focused methods mask attribute-related features in representations to protect sensitive information, but degrade recommendation accuracy. In contrast, performance-focused methods preserve accuracy by retaining these features but risk privacy leakage through uploaded representations. To balance privacy and performance, we propose PARD, a privacy-aware representation decoupling framework that explicitly decomposes representations into \textit{privacy-relevant} and \textit{privacy-irrelevant} components. Only the privacy-irrelevant part is uploaded to the server, and the privacy-relevant part is retained locally. We leverage Mutual Information (MI) objectives to facilitate this decoupling: (1) suppressing sensitive leakage by minimizing the MI between privacy-irrelevant representations and attributes, and (2) retaining attribute signals by maximizing the MI for privacy-relevant representations. Since exact MI computation is intractable, we derive variational bounds and estimate them using privacy estimators under adversarial and cooperative training paradigms. Experimental results demonstrate that PARD outperforms state-of-the-art methods in both recommendation accuracy and privacy preservation. The code is available at https://github.com/XuHao-bit/PARD.

Federated recommendation systems commonly protect user privacy by keeping user parameters on local devices, while exchanging item parameters for collaborative model training. However, such item parameters usually model items independently and suffer from both efficiency and effectiveness challenges, making communication costs grow with the item space and limiting cross-item generalization and robustness to noisy feedback. To address these limitations, we propose to model items via shared latent feature embeddings for communication. Residual Quantization (RQ) provides a natural way to instantiate this communication by representing each item with a short sequence of discrete code IDs, i.e., Semantic IDs (SIDs). However, directly applying centralized and static RQ-based recommendation to federated learning is non-trivial due to 1) private and biased historical interactions and 2) evolving collaborative information. We propose RQFedRec, an RQ-based federated recommendation framework for dynamic feature-embedding communication. To construct globally aligned codebooks without accessing private interactions, RQFedRec introduces an information distillation module. Each client first learns item embeddings that encode local collaborative information from private interactions, and then distills such information into feature-indexed codebooks under globally shared SIDs, making sparse and biased local signals more compatible with server aggregation. To adapt to evolving collaborative information, RQFedRec introduces a self-refining SID update module that dynamically refines global SID assignments from aggregated codebooks. Extensive experiments demonstrate that RQFedRec improves recommendation performance and reduces communication costs without relying on semantic information, while further benefiting from public semantics when available. Code is available at https://github.com/Mingzhe-Han/RQFedRec.

Federated recommendation (FedRec) suffers from sparse and highly heterogeneous interaction data, which can drive client-side item representations toward inconsistent semantic directions and cause feature conflict during aggregation. Addressing feature conflict in FedRec is non-trivial, as implicit feedback reflects user-specific preferences rather than stable item semantics, making it difficult to support prototype-based alignment as in standard FL. To address this issue, we propose MoSA, a FedRec-specific feature alignment framework that introduces multi-modal semantic anchors via item clustering. MoSA uses the resulting cluster assignments as globally consistent pseudo-labels and performs dual-level alignment: (i) a cluster-aware item representation alignment module to guide client-side item embedding matrix under partial item coverage, and (ii) a recommender modulation module to improve cross-client parameter compatibility while preserving personalization. Extensive experiments on real-world datasets demonstrate that MoSA consistently improves recommendation performance across representative federated recommendation backbones, and ablation studies validate the effectiveness of each component. \textit{\href{https://github.com/SunJY1116/MoSA_Code.git}{The codes are available here.}}

Federated recommendation is an emerging paradigm that enables collaborative model training while preserving user privacy by keeping raw interaction data on local devices. However, in this setting, each client performs local training solely on a single user's limited historical interactions, which further exacerbates the data sparsity issue that has long been a bottleneck limiting recommendation performance. In this paper, we delve into the data sparsity problem in federated recommendation and propose FedPerGen, a personalized generative augmentation framework for federated recommendation. Specifically, each client is required to upload a user preference proxy to the server. The server then feeds this proxy along with diversity-constrained random noise into a generative network to synthesize personalized pseudo-item embeddings for each client. To ensure the informativeness and diversity of the generated embeddings, we design three complementary optimization objectives: feature-contrastive loss, model-alignment loss, and embedding-diversity loss. Furthermore, a two-stage warm-up training strategy is proposed to guarantee reliable user preference proxies and high-quality pseudo embeddings. Extensive experiments on real-world datasets demonstrate that FedPerGen consistently improves various federated recommendation baselines. The code is available at https://github.com/xcheng202202/FedPerGen .

Federated recommendation enables collaborative model training while keeping user interaction data on local clients. A central problem in federated recommendation is how to aggregate useful information across clients for personalized recommendation. Existing personalized aggregation methods usually construct client relations from predefined parameter-based assumptions, such as parameter similarity or complementarity, and use these relations to determine aggregation weights. However, such methods construct a single global relation, which is insufficient to capture the hierarchical and multi-granularity nature of user relations in recommendation. Moreover, these predefined relations cannot directly reflect whether the related clients can improve prediction performance after aggregation. To address these limitations, we propose FedHUR, a federated recommendation framework for learning hierarchical utility-guided client relations. FedHUR takes item-item filters as the object for relation construction and aggregation. Specifically, it first aggregates and clusters each client's local information to obtain global hierarchical information. Each client computes hierarchical utility signals based on its local information and the global hierarchical information, indicating which collaborative information is useful for improving its prediction. The server uses these utility signals to retrieve clients that are useful to that client for further personalized aggregation. Extensive experiments on five real-world datasets show that FedHUR consistently outperforms existing federated recommendation baselines, demonstrating the effectiveness of hierarchical utility-guided client relation learning. Code is available at https://github.com/Mingzhe-Han/FedHUR.

Session-based Recommendation (SBR) achieves high performance in traditional centralized settings by using global item transition graphs to capture user intent. However, Federated Learning (FL) mandates collaborative training on isolated data, which makes the construction of such global structures unattainable. We identify this critical dilemma as ``Structure Collapse'', where the global graph remains fragmented despite model averaging. This collapse induces a ``Semantic Chasm,'' resulting in incompatible knowledge fusion and imprecise intent prediction. To break this impasse, we propose a novel Cross-Silo Federated Session-based Recommendation framework (CSFSR), which reconstructs global connectivity by harmonizing knowledge priors with local intents. Specifically, our Knowledge-Aware Initialization counteracts Structure Collapse by synthesizing shared attributes into a unified semantic space that anchors isolated clients. Further addressing the Semantic Chasm, the Graph-Guided Aggregation mechanism uses client signatures to direct personalized fusion, effectively filtering noise and ensuring robustness. Extensive experiments on three real-world datasets demonstrate that CSFSR significantly outperforms state-of-the-art federated baselines and exhibits superior noise robustness. Our code is available at https://github.com/Is-Jiejie/CSFSR.

Cross-domain recommendation (CDR) transfers preference knowledge across related domains, but federated deployment makes cross-domain alignment difficult because the behavioral anchors that align item spaces, such as overlapping users and shared interaction signals, are often sparse, unavailable, or privacy-sensitive across clients. To address this tension, we revisit federated CDR as generation over a stable semantic item language. By representing items as discrete semantic ID (SID) sequences derived from public item-side metadata, cross-domain item alignment is induced by a shared vocabulary rather than by exchanging private interactions or aligning domain-specific embeddings. Directly federating SID-based generators, however, introduces two design constraints: the SID tokenizer must remain fixed to preserve cross-client token consistency, which creates a semantic-only bottleneck because local collaborative filtering (CF) signals cannot be globally shared or aligned; meanwhile, standard federated averaging can cause negative transfer under domain heterogeneity. To overcome these constraints, we propose FedCGR, a federated generative CDR framework that keeps the item language stable and makes adaptation explicit. FedCGR injects local CF evidence through a reliability-aware semantic interface and trains a prototype-personalized generator that selectively aggregates shared parameters according to domain relatedness while keeping domain-specific quantities local. Experiments on six Amazon cross-domain scenarios show that FedCGR consistently outperforms federated generative baselines and achieves competitive performance against strong sequential and federated CDR methods under both full-ranking and sampled evaluation protocols. Our code is publicly available at https://github.com/ZhuodLiu/FedCGR.

Cross-domain recommendation (CDR) boosts personalization by transferring knowledge across domains, but its privacy implications remain largely overlooked. In this paper, we uncover a critical threat in CDR: cross-domain user re-identification can be performed in realistic black-box settings using only the top-$K$ recommended items returned by deployed systems. We propose a list-supervised shadow-model attack that treats the target model's top-$K$ lists as implicit feedback and mimics its ranking behavior through a pairwise ranking objective. To amplify identity signals that persist across domains---even with disjoint item sets---the attacker learns disentangled user interests, separating domain-specific preferences from domain-invariant representations through contrastive constraints. The resulting representations support an effective classifier for determining whether two cross-domain accounts belong to the same individual. Extensive experiments across multiple CDR settings show that re-identification leakage is pervasive in modern models. Finally, we present practical defense mechanisms that mitigate linkability while preserving recommendation utility, highlighting the need for privacy-aware CDR design. \textit{\href{https://github.com/SleepingZSJ/LiSA-Link.git}{The codes are available here.}}

Query correction is a critical entry point in modern search pipelines, demanding high accuracy strictly within stringent latency constraints. Chain-of-Thought (CoT) reasoning improves accuracy but introduces excessive latency that limits real-time deployment. A natural alternative is to output an answer before reasoning to reduce latency; however, under autoregressive decoding, such an answer-first paradigm suffers from a \emph{decoupling problem}: the early answer is generated independently of the subsequent reasoning, and therefore cannot benefit from it. To address this issue, we propose Sandwich Reasoning (SandwichR), a novel approach that explicitly aligns a fast initial answer with post-hoc reasoning, enabling low-latency query correction without sacrificing reasoning-aware accuracy. SandwichR follows an ``Answer–Reasoning–Answer'' paradigm, producing an initial correction, an explicit reasoning process, and a final refined correction. To align the initial answer with post-reasoning insights, we design a consistency-aware reinforcement learning (RL) strategy: a dedicated consistency reward enforces alignment between the initial and final corrections, while margin-based rejection sampling prioritizes borderline samples where reasoning drives the most impactful corrective gains. Additionally, unlike existing datasets that primarily focus on simple spelling errors, we construct a high-quality query correction dataset tailored to semantically sensitive errors. Experiments on three real-world domains show that SandwichR delivers the best accuracy-latency trade-off among evaluated methods at comparable model scale, achieving accuracy on par with standard CoT while reducing inference latency by 40--70\%, and offering a promising solution for latency-sensitive applications.

LLM-generated search queries are widely used to augment IR evaluation, yet they may contain concepts that presuppose answer-side document knowledge, violating the information-access boundary of pre-search users. Existing validation metrics---overlap, diversity, effectiveness---cannot distinguish rare human-tail variation from candidate answer-side intrusion. We introduce \emph{concept provenance}, a framework that assigns query concepts to backstory-supported, human-central, human-tail, and candidate answer-side zones, operationalizing a boundary that retrieval metrics alone cannot detect. Applying concept provenance to 77,004 queries across 100 UQV100 topics, 8 LLMs, and 5 prompt conditions with two extraction pipelines (cross-pipeline token-HCIR Spearman $\rho = 1.0$ over five condition means), we find candidate answer-side concepts constitute 7.40\% of non-generic concepts, appearing in 97 of 100 topics with topic explaining ${\approx}67\%$ of variance. Human validation yields 68.2\% relaxed precision, revealing a dual mechanism: knowledge intrusion (45.5\%) and deployment intrusion (45.0\%). Diagnostic probes show these concepts carry disproportionate localized retrieval effects (deletion $d = -0.47$, exceeding random deletion $d = -0.34$) but explain less than 2\% of aggregate evaluation variance, positioning concept provenance as a boundary-compliance diagnostic rather than an evaluation-shift predictor. Under the tested conditions, no prompt condition eliminates intrusion; post-generation concept-provenance selection achieves 99\% elimination.

Large language models (LLMs) are increasingly used in information retrieval (IR) pipelines as relevance judges and re-rankers. Yet most analyses remain output-centric, evaluating generated labels or scores while offering limited insight into how relevance is represented inside the model. In this work, we study whether query-document (q-d) relevance is linearly decodable from residual-stream activations in instruction-tuned LLMs, how this internal signal compares with generated relevance judgments, and whether it transfers across languages. Using the TREC DL20 and MIRACL test collections, we guide medium-scale LLMs (4-9B parameters) with UMBRELA-style relevance judgment prompts, extract last-token activations from every transformer layer, and train linear probes to predict relevance labels. We compare probe predictions with generated judgments and use TREC DL20 to test whether probe-derived pseudo-labels preserve system rankings when compared with human judgments. Our results suggest that q-d relevance is encoded as a depth-dependent internal signal: probe performance is weak in early layers and strongest in middle-to-late layers, indicating that relevance becomes more linearly accessible after contextual integration. Most importantly, in several models, validation-selected probes match or outperform generated judgments and better preserve system rankings, revealing a separation between internal relevance representation and external expression. Multilingual experiments further suggest partial cross-language portability, although transfer remains weaker than within-language decoding. Overall, this work provides a representation-level perspective on LLM-based relevance assessment. Layer-wise probing can help diagnose where relevance emerges, when generated judgments fail to reflect internally available evidence, and how relevance representations vary across languages, datasets, and model families.

Web search, product search, and question-answering retrieval systems often assign a relevance label and confidence score to each query-candidate pair. The relevance label describes how well a page, product, or passage matches the query, while the confidence often guides downstream use or fallback decisions. Post-hoc calibration is therefore needed because misaligned confidence can make systems over-trust wrong predictions or unnecessarily defer correct ones. However, calibration mainly aligns confidence with average correctness, and does not remove predicted-label-dependent reliability differences that remain within the same calibrated confidence level. We address this gap with Label-wise Monotone Reliability Projection (MRP), which learns label-wise monotone functions that map calibrated confidence to correctness reliability while preserving the original predicted labels and class probabilities. The resulting reliability score reranks fixed predictions according to residual risk. Across six information access relevance datasets and multiple post-hoc calibrators, MRP improves reliability reranking and average fallback utility while preserving full-coverage accuracy and ECE. Structural ablations show that the main gains come from label-wise residual reliability rather than from global confidence remapping. We further analyze when MRP reliability scores can be embedded back into top-label probability geometry, showing that this projection is useful as a compatibility analysis but is distinct from the main reliability-reranking objective. The implementation is available at https://github.com/inu0104/MRP.

Static pruning is widely used to accelerate sparse neural retrieval, yet existing studies each validate their conclusions within a single custom pipeline, leaving it unclear which findings transfer to modern engines with different index organizations and dynamic pruning mechanisms. We present the first cross-engine pruning portability study, evaluating static pruning strategies across three engines—a controlled C++ pipeline (exhaustive inverted index), BMP (block-max pruning), and SEISMIC (clustered inverted indexes)—on two benchmarks (MS MARCO, Natural Questions) with two encoders spanning opposite query-density regimes (SPLADE: 44 avg. query terms; V3-GTE: 7 avg. query terms), totaling 1,140 experimental configurations, with an additional deep-judgment validation on TREC DL 2019/2020. We find that index-side pruning (document and posting-list) is portable: it consistently reduces latency (1.2–6.6×) and index size (18–82%) across all engines because sparse retrieval is memory-bound—a conclusion we support with cache-miss, TLB, and IPC profiling. In contrast, query pruning is already internalized by modern engines: it yields 4–11× speedup on the exhaustive pipeline but is subsumed by BMP's ? and SEISMIC's query_cut. Static pruning complements dynamic pruning: on BMP, combining document and query reduction yields 2.5× speedup with NDCG@10 within 0.003 of the exact baseline. Finally, NDCG@10 saturates while Recall@10 is still in the ~85–95% range across all three engines, providing a portable stopping criterion: practitioners can push pruning to this knee without visible ranking degradation. Together, these findings answer what transfers (index-side pruning), what breaks (query pruning), and what still helps (static atop dynamic pruning). Code is available at https://github.com/zirui-song-18/cross_engine_static_pruning.

Generative retrieval models perform document retrieval by autoregressively generating document identifiers (DocIDs). This process naturally forms a sequential decision problem, i.e., the model makes a sequence of token-level decisions, selecting a DocID token at each decoding step, with the resulting complete sequence identifying the retrieved document. However, relevance feedback is available only after the complete DocID has been generated and mapped to a document, resulting in a granularity mismatch between token-level generation decisions and document-level retrieval supervision. Consequently, existing reinforcement learning methods for generative retrieval rely on sequence-level rewards, assigning the same document-level relevance signal to every decoding step. Such uniform credit assignment obscures the contribution of individual token decisions, making it difficult to identify which decisions contribute to retrieval success or failure. In this paper, we propose Token-Level Credit Assignment for Generative Retrieval (TCA), a fine-grained reinforcement learning framework that aligns the granularity of credit assignment with that of autoregressive DocID generation. Unlike assigning a single reward to an entire generated DocID, TCA derives fine-grained rewards by comparing the hidden-state trajectory of each generated DocID with the gold DocID trajectory obtained from a frozen reference model. These trajectory-based rewards provide differentiated feedback across decoding steps, allowing the policy to reinforce generation paths that remain aligned with the target DocID. Moreover, TCA decouples token-level credit assignment from policy optimization and can be instantiated with both GRPO and PPO. Experiments on MS MARCO and NQ benchmarks show that our method consistently outperforms baselines, demonstrating the effectiveness of fine-grained supervision for aligning autoregressive DocID generation with retrieval objectives.

Dense retrieval has become a cornerstone of modern local-lifestyle e-commerce search by encoding queries and items into semantic embedding spaces. While recent advancements have transitioned from BERT-based embedding models to Large Language Models (LLMs), most approaches still treat LLMs as static text encoders, neglecting their inherent reasoning capabilities. Furthermore, standard dense retrieval models remain query-centric, which is insufficient in e-commerce scenarios where sparse and ambiguous queries create an intent gap that can only be bridged by the rich context of user history. Meanwhile, existing personalized retrieval methods typically rely on implicit embedding interactions, which lack the reasoning capability to effectively disambiguate user intent from noisy historical behaviors. To address these challenges, we propose Think-to-Personalize (TTP), a novel framework that unifies explicit user-centric intent reasoning with dense retrieval. By reasoning over the user's historical purchase sequence, TTP explicitly deduces latent personalized needs and generates an intent-enhanced query, which is then encoded into a unified dense embedding. Specifically, we design a two-stage training paradigm: (1) A Supervised Fine-Tuning (SFT) stage that establishes cold-start capabilities; (2) A Reinforcement Learning (RL) stage that aligns the reasoning process with retrieval utility using Group Relative Policy Optimization (GRPO). Extensive experiments on both proprietary and public benchmarks demonstrate that TTP significantly outperforms state-of-the-art baselines. Furthermore, in online A/B tests, it achieved a +0.46% lift in order volume, validating its practical effectiveness and establishing a new paradigm for reasoning-driven personalized dense retrieval.

Production search indices operate under strict storage and latency budgets that prevent indexing the full corpus (e.g.: web documents). When only a fraction of documents can be retained, the system must decide which ones to keep. Existing pruning methods make this decision using query-agnostic signals, including lexical quality, embedding magnitude, and corpus centrality or single-feature query-aware heuristics (e.g., query similarity, nearest-neighbour coverage), without optimising which documents actually contribute to answering the queries the system will serve. We recast index pruning as query-aware subset selection: given a fixed budget and an expected query workload, the goal is to retain the document subset that maximises retrieval performance for the queries the system will serve. We introduce Query-aware Document Selection (QDS), a lightweight framework that scores documents using 13 retrieval-structural features derived from query–document interactions and index redundancy, capturing how many queries a document serves, whether viable alternatives exist, and how replaceable the document is within the retained set. A linear scorer trained with Evolution Strategies optimises nDCG@10 of the pruned index end-to-end, requiring no differentiable surrogate while remaining fully interpretable. % Across 10 IR benchmarks (1K–120M documents), four dense encoders plus BM25, and three modalities (text, image, audio), QDS consistently outperforms query-agnostic and single-feature query-aware baselines under aggressive pruning budgets. We show that coverage-oriented heuristics degrade at web scale by over-selecting high-frequency hard negatives, i.e., documents that are nearest neighbours to many queries but relevant to none. Given its simplicity, QDS can operate on existing indexes of 100M-1B documents in minutes with limited hardware.

Clickbait headlines are purposefully designed to omit crucial news information, thus evoking strong curiosity and enticing readers to click. Naturally, such heightened curiosity is typically accompanied by more diverse imaginings of the news content and stronger emotional arousal among readers. Inspired by these cognitive and affective reactions, we develop a novel clickbait detection method, entitled \textbf{C}lickbait Detector with \textbf{A}ugmentation of \textbf{C}ognitive \textbf{I}magining and Emotion \textbf{A}rousal (\textbf{\baby}). Specifically, a persona-induced multi-agent system (MAS), which simulates real-world news readers, is exploited to generate multiple content imaginings of a given headline. Meanwhile, we suggest a dimensional emotion recognition (DER) task within a multi-task learning framework to estimate the emotional arousal elicited by the headline. Accordingly, clickbait headlines tend to induce more diverse imaginings and stronger emotional arousal than non-clickbait ones. These relative differences are formulated as additional training objectives, thus enabling a more effective capturing of curiosity semantics in clickbait headlines. Finally, \baby derives discriminative clickbait features augmented by human-like imaginings and emotion knowledge. Extensive experiments on benchmark datasets demonstrate that \baby consistently outperforms strong baseline methods.

Online platforms are increasingly shaping political communication, influencing how narratives emerge and gain traction. Unfortunately, this has led to an acceleration in the spread of false information. One way to analyze misinformation dynamics is to measure whether public discussions support or refute verified claims. However, most existing stance detection methods struggle to keep up with novel claims without retraining on new labeled data. In this paper, we extend claim-based stance detection research, explicitly focusing on generalization to unseen claim targets and model performance in-the-wild. We introduce a deep learning architecture designed to model the relational interactions between a claim and a post by combining a two-way cross-attention mechanism and supervised contrastive learning to learn stance features that transfer to unseen claims. Intuitively, this approach combines both instance-level reasoning associated with claim and post tokens and higher order reasoning associated with the stance detection task. Our experiments on benchmark datasets demonstrate improved generalization performance across multiple domains when compared to state-of-the-art methods. Finally, we conduct a case study that uses our proposed model to measure the stance of voting claims on social media from the 2024 U.S. election, highlighting variations in misinformation across different voting-related themes and the benefit of our model for timely and scalable analysis of online discourse.

As large language models advance, news increasingly contains both human-written and LLM-generated content. Existing models tend to exploit linguistic fingerprints in LLM-generated texts rather than deep semantics for veracity inference, leading to shortcut learning and detection bias when classifying both human-written and LLM-generated news. To address this issue, we propose a \textbf{S}emantic \textbf{D}ecoupling and \textbf{D}ynamic \textbf{E}nhancement (\textbf{SDDE}) plugin, which can be integrated into fake news detection models as a plug-and-play module to mitigate linguistic fingerprint interference and enhance transferable semantic information. Specifically, SDDE first obtains original text representations from the backbone model and denoises them to extract purified features. Subsequently, based on these features, adversarial training and auxiliary supervision are applied to reduce the influence of linguistic fingerprints while preserving deep semantics. Meanwhile, a cross-attention mechanism supplements semantic details of the purified features to obtain enhanced features. A token-level adaptive gating dynamically fuses enhanced features with the original text representations. Finally, the fused representations are reinjected into the backbone model for training. Experiments show that incorporating the SDDE plugin yields absolute accuracy improvements of 1.19\%--6.29\% across backbone models on two datasets, demonstrating its effectiveness and generalizability. The code is available at \url{https://github.com/cgao-comp/SDDE}.

Effective multimodal fake news detection necessitates considering extracted cues not merely as features for passive aggregation but as distinct evidence for selective integration. Nevertheless, current methodologies primarily depend on static fusion architectures, lacking the ability to dynamically adjust to the particular evidence needed for each given sample. The constraint arises from their inability to effectively model the hybrid decision space: routing to specific evidence branches while regulating their respective contributions. To bridge this gap, we propose PAGE, a Policy-driven Adaptive Gating framework utilising Evidence fusion that explicitly decouples multimodal features into complementary content and editing evidence. To tackle the optimisation constraints of this hybrid action space, we formulate evidence fusion as a policy learning problem optimised via reinforcement learning. Specifically, an adaptive policy gate issues a hybrid action for each instance: it performs a discrete routing decision to establish the optimal fusion structure, while simultaneously predicting continuous gating values to calibrate branch interactions. Guided by task-level detection rewards, the framework jointly optimises these coupled decisions. Extensive experiments on benchmark datasets demonstrate that PAGE achieves satisfactory performance, significantly improving detection accuracy over state-of-the-art baselines. The source code is available at \url{https://github.com/cgao-comp/PAGE}.

Multimodal fact-checking systems are increasingly expected to return not only a veracity label but also a citation-grounded re port. Yet a correct label may still be accompanied by query drift, hallucinated citations, unsupported evidence statements, or over strong conclusions. We argue that many such failures arise from evidence-state management errors in two local windows: retrieval plan formation before evidence retrieval and citation integration after evidence retrieval. We propose GESR, a training-free inference time framework that uses an executable claim-local evidence-state graph to constrain local repair actions under a fixed retrieval budget with one real retrieval call per example. Unlike graph-as-context prompting, GESR does not merely expose graph information to the generator. It uses graph checks to accept or reject typed actions, including entity, time, source, and modality edits before retrieval, and citation replacement, claim weakening, modality-support inser tion, and verdict downgrading after retrieval. To evaluate citation grounding, we build MR2-E, a controlled evaluation protocol on top of MR2 with a frozen retrieval index, opaque citation identifiers, top-5 citation resolution, and sentence-level support judgments. On MR2-E under the fixed single-retrieval-call setting, GESR shows statistically inconclusive label-level F1 differences but larger and more stable improvements in evidence consistency, citation accu racy, and human-judged citation grounding than verifier-repair and graph-exposure baselines.

Automatic fact-checking systems assess the veracity of claims given evidence from relevant documents. Large Language Models (LLMs) have demonstrated strong performance in fact-checking due to their general reasoning capabilities. However, it remains unclear whether they faithfully make use of the evidence provided to reach veracity judgments or rely on parametric knowledge. To investigate this, we introduce Fact-Ablated Evaluation (FAE), a new evaluation framework that iteratively ablates the cited evidence to assess whether LLMs revise their predictions accordingly. Our empirical results show that current off-the-shelf LLMs as fact-checking systems rely more on their parametric knowledge than on the evidence provided. To bridge this gap between prediction accuracy and evidence grounding, we propose REAL (Rigorous Evidence Ablation Learning), a training framework that promotes evidence-dependent verification through counterfactual evidence supervision for the LLM-as-verifier models. Experiments on four fact-checking datasets across different domains demonstrate that models trained with REAL obtain superior evidence-dependent capabilities compared to standard fine-tuned models. Our findings highlight that strong fact-checking performance can still coexist with weak evidence dependency, while REAL encourages veracity predictions to remain more closely tied to the availability of supporting evidence.

Misinformation diffusion at scale on social media poses serious societal risks, as it can distort collective decision-making, undermine public health interventions, bias political discourse, and trigger large-scale behavioral cascades during emergencies. Most existing approaches rely heavily on observable content signals or user-level features, primarily capturing surface-level correlations while overlooking the latent causal mechanisms governing diffusion dynamics. To bridge this gap, we formulate misinformation diffusion prediction from a causal perspective and propose MisTran, a mechanism-aware framework that models diffusion through causal mediation. Specifically, MisTran introduces a structured mediator to represent propagation mechanisms that connect persuasive content to cascade growth, enabling the model to move beyond correlation-based inference toward mechanism-oriented prediction. To further improve robustness, we incorporate LLM-based counterfactual reasoning to encourage causal consistency during training. By constructing hypothetical variations of posts, the model learns invariant propagation patterns that reflect stable causal structure rather than spurious correlations. Extensive experiments on a real-world misinformation diffusion benchmark demonstrate that MisTran consistently outperforms strong baselines. Ablation and interpretability analyses further validate the effectiveness of mechanism-constrained modeling and provide interpretable insights into the dynamics of misinformation diffusion.

AI-generated images increasingly enter information-access systems where provenance signals inform search ranking, moderation, and news verification, often only after web and social-media redistribution. The practical task is therefore metadata-free visual content provenance rather than pristine-image classification: recompression, reposting, resizing, and re-digitization can alter the evidence available to a detector. Existing detectors typically rely on either low-level artifact cues or high-level representation cues, but these cues exhibit asymmetric reliability across the content life cycle. We present HybridDet, a representation-first dual-expert framework for redistribution-robust AI-generated image provenance. HybridDet combines Q-Table Transfer (QTT), which aligns real and synthetic samples through empirical JPEG quantization-table transfer during training, with Patch2DCT, which converts multi-scale DCT-band statistics into ViT-compatible patch tokens. Instead of dense fusion, HybridDet adopts a Confidence-Aware Router (CAR) that keeps QTT as the default path and invokes Patch2DCT only under validation-calibrated selective routing, yielding a validation-calibrated selective provenance signal. Experiments on ten clean-transfer and redistribution-oriented benchmarks show that HybridDet improves the robustness \& generalization balance for binary AI-generated visual provenance, especially on real-world redistribution and recent text-to-image generators. Ablations, routed-subset analyses, calibration metrics, selective-decision analysis, and real-photo checks show that HybridDet selectively exploits complementary provenance evidence under distribution shift. Code is available at \url{https://github.com/graydove/HybridDet}.

Recent large reasoning models (LRMs) have achieved strong performance on complex reasoning tasks by generating a long chain-of-thought (Long-CoT). However, such lengthy reasoning is often unnecessary for simple queries, leading to additional computation and latency. Existing approaches to adaptive reasoning typically rely on retraining the model or designing sophisticated prompting, which are either prohibitively expensive or highly sensitive to the prompt formulation. Model merging provides a more balanced alternative for adaptive reasoning by avoiding expensive training and integrating Long-CoT and Short-CoT behaviors. However, existing merging methods are often static and input-agnostic, or rely on costly all-layer calibration, which limits their effectiveness for query-adaptive reasoning. To tackle these challenges, we propose Query-adaptive Layer Selective Merging (QA-Merging), an activation-based merging framework that integrates a Long-CoT model and a Short-CoT model to obtain a query-adaptive reasoner without training from scratch or requiring large-scale additional data. QA-Merging first constructs a small pattern-labeled calibration set that assigns each query an appropriate reasoning pattern. Motivated by our empirical analysis that Long-CoT and Short-CoT behaviors diverge unevenly across transformer layers, QA-Merging identifies layers with high reasoning pattern divergence and calibrates only these layers through feature alignment and contrastive shaping, while applying closed-form hidden-state correction to the remaining layers. Experiments on seven widely used reasoning benchmarks across two model scales demonstrate that QA-Merging reduces inference cost and maintains strong performance. The code is available at https://anonymous.4open.science/r/QA-Merging-q1a2.

Chain-of-Thought (CoT) reasoning, with long or short variants, has become a central paradigm for enhancing the reasoning ability of large language models. Long CoT aims at achieving high accuracy in complex multi-step reasoning at the cost of high inference overhead, while short CoT focus on offering a more efficient but less capable alternative. To balance reasoning accuracy and reasoning efficiency, collaborative approaches combining long and short CoT have been widely explored. However, existing collaborative approaches either operate token-level pruning, risking disruption to reasoning coherence, or rely on instance-level routing, which over-look the excessive self-verification within long CoT reasoning. To address these limitations, a \underline{sem}antic step-level reasoning optimization approach for collaborative \underline{CoT}s, named SemCoT, is investigated in this paper. Technically, verification redundancy and trigger mechanisms in long CoT reasoning are first analyzed. Specifically, a dynamic trigger leveraging lexical cues and logit uncertainty is designed to detect the onset of self-verification in long CoT reasoning. A finetuned short CoT model is subsequently employed to filter trivial verification, with redundant cases truncated through context steering. Confident phrases are finally embedded to ensure contextual coherence. Experimental results demonstrate that SemCoT reduces average reasoning length by 48\% across several benchmarks while maintaining similar accuracy, highlighting its practical effectiveness.

Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks, but their long chain-of-thought trajectories incur substantial inference latency and computational cost. Small-large model collaborative reasoning mitigates this issue by offloading part of the reasoning steps to a small language model (SLM) while preserving large language model (LLM) computation for more demanding steps. Existing draft-and-verify methods use an external verifier to reject low-scoring SLM steps and regenerate them with the LLM, introducing per-step verification overhead. Moreover, we observe that many LLM rewrites remain semantically close to the rejected SLM drafts, causing unnecessary decoding cost. Further analysis shows that these redundant rewrites often occur on local derivation steps, whose content is largely constrained by recent context. Motivated by these observations, we propose \textbf{CriticRouter}, a verifier-free step-wise routing framework that reduces redundancy based on structural criticality as a routing signal. Specifically, structural criticality measures how sensitive the SLM's next-step generation is to perturbations in earlier reasoning history, and low-criticality steps are viewed as locally stable and delegated to the SLM, whereas high-criticality steps are treated as history-sensitive and routed to the LLM. To reduce probing overhead, CriticRouter further adopts short concurrent probes and progressive mask advancement for better cache reuse. Experiments on multiple benchmarks show that CriticRouter achieves comparable accuracy to standard LLM reasoning while significantly reducing latency, demonstrating a better utility-efficiency trade-off.

The rapid adoption of large language models (LLMs) has created demand for faster responses and lower costs. Semantic caching, reusing semantically similar requests via their embeddings, addresses this need but breaks classic cache assumptions and raises new challenges. In this paper, we explore offline policies for semantic caching, proving that implementing an optimal offline policy is NP-hard, and propose several polynomial-time heuristics. We also present online semantic aware cache policies that combine recency, frequency, and locality. Evaluations on diverse datasets show that while frequency based policies are strong baselines, our novel variant improves semantic accuracy. Our findings reveal effective strategies for current systems and highlight substantial headroom for future innovation. All of our code is open source.

As large language models increasingly support long-context inference, the computation and memory access overhead of attention and KV caches have become critical bottlenecks for inference latency and system scalability. Existing approaches mitigate these costs through KV cache dropping, approximate attention, or system-level optimizations, but often incur non-trivial trade-offs between efficiency, accuracy, and implementation complexity. In this work, we propose LSH-KV, a coarse-to-fine KV retrieval framework for efficient long-context LLM inference. LSH-KV uses locality-sensitive hashing (LSH) as a lightweight, training-free coarse partitioning layer for similarity-preserving organization of global KV entries. Within each hash bucket, LSH-KV further uses product quantization with offline-precomputed codebooks for efficient approximate similarity estimation, followed by exact re-ranking to preserve attention accuracy. To further improve system efficiency, LSH-KV introduces a score-aware block-level cache management strategy that aligns KV movement between CPU and GPU with retrieval relevance. Experimental results on long-context benchmarks demonstrate that LSH-KV significantly reduces end-to-end inference latency while preserving generation quality, even when selective attention is performed using only a fraction of the input tokens. These results show that LSH-KV enables scalable, robust, and accurate inference under long-context settings. The code is available at https://github.com/HarryFunn/lsh-kv.

Long system prompts are widely used to steer Large Language Models (LLMs), but repeatedly processing them at inference time is inefficient and consumes valuable context budget. This motivates a central question: can the behavioral effect of a long system prompt be retained using only a minimal learned representation? To enable this, we propose a lightweight training framework that learns a single Behavior-Equivalent Token ([BE]). The framework first trains [BE] to encode the semantic content of the original system prompt via reconstruction, and then distills the prompt's downstream behavior into this single token. Importantly, our method requires no update to the pretrained LLM weights, no auxiliary compression models, and no labeled responses. Empirical evaluations on three datasets show that replacing long prompts with a single [BE] token yields up to a 3000x prompt compression ratio, while retaining about 98% of the downstream performance of the original system prompts. This substantially reduces inference cost and frees nearly the entire context window for user inputs and model outputs. The implementation code is publicly available for reproduction.

The massive redundancy of visual tokens in Vision-Language Models (VLMs) creates a critical computational bottleneck during large language model (LLM) prefill. Recent training-free pruning methods provide plug-and-play acceleration, yet their pruning decisions often rely on limited visual states or delayed in-LLM signals, which may underexploit both the layer-wise evolution of visual representations and the query-specific relevance of local regions. We propose DATP (Depth-Aware Token Pruning), a training-free early-pruning framework that selects compact visual tokens before they enter the LLM. DATP combines two complementary signals: Depth-aware Layer Aggregation (DALA), which aggregates multi-layer visual saliency trajectories to construct a high-recall structural prior, and Cross-modal Attention Gating (CMAG), which introduces a lightweight prompt-conditioned relevance signal to suppress visually salient but task-irrelevant tokens. Experiments on LLaVA-v1.5, LLaVA-NeXT, and Qwen2.5-VL show that DATP achieves a favorable accuracy-efficiency trade-off under aggressive token budgets. On LLaVA-v1.5, DATP preserves 95.7% of the full-token performance while retaining only 11% of the visual tokens. Across the reported multimodal reasoning and hallucination-sensitive benchmark settings, DATP generally outperforms the evaluated training-free pruning baselines.

Large vision-language models (LVLMs) have recently been extended to handle multiple input images, but this extension dramatically increases the visual token budget, making visual-token pruning not merely an efficiency option but a practical necessity. Existing visual-token pruning methods, primarily designed for the single-image setting, rely on attention-derived saliency signals to estimate per-token importance. However, this attention-centric design transfers poorly to the multi-image regime: it causes substantial performance degradation on multi-image benchmarks, suffers from strong positional bias, and loses discriminative attention signals as the number of input images increases. To test whether attention-selected tokens are truly important for multi-image reasoning, we conduct two complementary causal probes—Leave-One-In and Leave-One-Deletion—which measure token importance by restoring individual tokens to a zeroed visual sequence or removing them from the full sequence. Attention-based pruners provide only marginal gains over random selection, showing that attention is a weak proxy for visual-token importance in multi-image reasoning. Motivated by this gap, we propose Relation and Information Aware Pruning (RI-Prune), a training-free, attention-free pre-decoder selector for multi-image visual token pruning. RI-Prune scores each visual token using a z-normalized combination of three signals computed from projector outputs: query relevance, cross-image correspondence with compact per-image prototypes, and intra-image representativeness. By integrating these signals, RI-Prune performs balanced per-image Top-$K$ selection to preserve information across input images. Across multiple multi-image benchmarks and three LVLM backbones, RI-Prune achieves a stronger trade-off between token reduction and task performance than prior pruning methods, particularly under aggressive token budgets.

Medical image diagnosis has achieved significant progress with deep learning, yet existing methods often rely on isolated visual evidence and lack the ability to effectively leverage similar cases and external knowledge. In clinical practice, diagnosis is typically supported by similar historical cases and their associated symptoms. To explicitly model this evidence-based diagnostic process, we propose MKG-CARE, a framework that performs case-aware reasoning using multimodal knowledge graphs for explainable medical image diagnosis. Specifically, we construct a case-aware multimodal knowledge graph as a structured diagnostic memory, where diseases, images, and symptoms are hierarchically organized. Given an input image, MKG-CARE adaptively retrieves similar cases from this memory and extracts their corresponding case-centered subgraphs. We further introduce a knowledge propagation and injection mechanism, where an image-centric Graph Attention Network aggregates heterogeneous semantics within the retrieved case subgraphs, followed by bidirectional cross-modal attention to align and inject the aggregated case knowledge into visual representations. To mitigate retrieval noise, we design a confidence-calibrated decision refinement scheme that estimates each retrieved case’s reliability from prediction confidence and sample similarity, and reweights its contribution to the final prediction for interpretable case-level evidence attribution. Extensive experiments on multiple medical imaging datasets demonstrate consistent improvements over strong baselines, while ablation and qualitative analyses validate the effectiveness and interpretability of our method. The code is available at https://github.com/lyxuan1022/MKG-CARE.

Radiology report generation (RRG) translates medical images into clinical reports, yet existing systems diverge from real radiological practice in three respects. First, they are formulated at the image level, consuming a single image or a fixed pair of current/prior images, and discarding the variable visual evidence within a patient study. Second, report generation flows through a single end-to-end pipeline. Even when auxiliary classifiers or retrieved priors are incorporated, they remain auxiliary inputs rather than independent sources of clinical judgment. Third, both retrieval and validation operate without a finding-level clinical anchor: retrieval falls back on image-feature similarity, while validation relies on prompt-level heuristics, leaving generated reports weakly grounded. We propose DAMEC, a Disease-Aware Multi-Expert Consensus framework that recasts RRG as study-level, clinical-context generation. Four heterogeneous experts process a variable number of images from a patient study and emit per-expert clinical signals, which a trainable disease-aware consensus module aggregates into a unified consensus. The consensus, together with disease states, uncertainty, longitudinal changes, and attribute modifiers, is externalized into a structured Clinical Findings (CF) descriptor, serving as the finding-level anchor for both clinical-context retrieval and post-generation validation. On MIMIC-CXR, MIMIC-ABN, Two-view CXR, and CheXpert Plus, DAMEC consistently improves both natural language generation and clinical efficacy metrics over strong baselines. Code is available at: https://github.com/ku-milab/DAMEC.

Domain generalization has become a critical challenge in predictive healthcare, where different patient groups exhibit shifting data distributions that degrade model performance. Still, regular domain generalization approaches often struggle in clinical settings due to (1) the absence of domain labels and (2) the lack of clinical insight integration. To address these challenges in healthcare, we aim to explore how medical ontologies can be used to discover dynamic yet hierarchy-grounded patient domains, a partitioning strategy that remains under-explored in prior work. Hence, we introduce UdonCare, a hierarchy-pruning method that iteratively divides patients into latent domains and retrieves domain-invariant (label) information from patient data. On public datasets (MIMIC-III, MIMIC-IV, and eICU), UdonCare shows superiority over eight generalization baselines across four representative clinical prediction tasks with substantial domain gaps, highlighting the potential of medical knowledge for enhancing model generalization.

Causal structure learning in clinical domains faces a persistent challenge: tabular datasets are often limited, while valuable domain knowledge exists primarily in unstructured clinical narratives. Recent work has explored large language models (LLMs) as sources of knowledge to refine causal graph discovery or define causal relations. However, relying directly on LLM-generated causal judgments can be problematic, as such outputs may reflect spurious associations, domain bias, or statistical patterns learned from text rather than reliable causal mechanisms. In this work, we propose a different perspective: instead of treating LLM outputs as definitive causal judgments, we use them as weak ordinal supervision integrated into the Hill-Climbing score through a soft regularization mechanism that remains correctable by observational data. Our framework extracts candidate causal relations from clinical narratives and assigns them ordinal plausibility labels using a domain-specialized LLM. We evaluate this approach on three established Bayesian network benchmarks (Alarm, Child, and Hepar II) and a real-world clinical dataset from cBioPortal, comparing Hill Climbing with and without ordinal priors across varying sample sizes. The framework consistently improves structural recovery, with statistically significant gains persisting even in high-data regimes. Overall, the results suggest that ordinal LLM supervision provides a robust inductive bias that helps resolve statistical ambiguities and mitigate structural errors in clinical causal discovery.

Offline sepsis treatment aims to learn treatment policies from retrospective ICU trajectories. Decision Transformer-based methods provide a natural sequence modeling framework for this task, but they mainly rely on raw physiological states and historical clinician actions. Such raw trajectory tokens may not explicitly capture clinical semantics that clinicians use in practice, such as patient severity, recent treatment response, and current risk priorities. To address this limitation, we propose \textbf{ClinSDT}, an LLM-assisted Decision Transformer framework for offline sepsis treatment. ClinSDT uses a frozen LLM as a semantic bridge rather than a treatment decision maker. It constructs three clinical semantic views from ICU trajectories, namely \textsc{Profile}, \textsc{Context}, and \textsc{Advisory}, and inserts their encoded representations as independent tokens into a Decision Transformer policy. To support policy refinement and rollout, ClinSDT further trains a Treatment-Response Forecaster to estimate the next physiological state, SAPS-II change, and predictive uncertainty from state--action inputs. In the second training stage, the forecaster screens local counterfactual treatment candidates, and only low-uncertainty candidates with lower predicted SAPS-II increase are distilled into the policy after semantic regeneration. During inference, ClinSDT performs closed-loop treatment rollout by refreshing semantic inputs from the currently available patient state. Experiments on MIMIC-III and MIMIC-IV show that ClinSDT achieves consistent relative improvements over representative imitation learning, offline reinforcement learning, and Decision Transformer baselines under the same rollout protocol. Code and data are available at \href{https://github.com/KAG778/ClinSDT}{ClinSDT}.

Medication recommendation aims to generate effective drug combinations based on longitudinal clinical records of patients, where historical prescriptions and relevant candidate medications serve as important references to improve performance. However, existing methods treat these medications as individual entities and overlook structured interactions among clinically co-occurring drugs that jointly contribute to treatment effects. In addition, medications lack intrinsic principles to guide the modeling of semantic compositions describing such interactions, making it challenging to capture coherent clinical structures and resulting in a gap between structured patient representations and flat medication references. To address these challenges, we propose PACRM, a novel medication recommendation framework that captures semantic compositions to summarize collective drug effects through multi-level structures. Specifically, the model introduces a GAT-based voting strategy to organize drugs working as combinations jointly into structured compositions, and further performs temporally constrained and dynamically determined cross-combination voting over historical prescriptions and selected candidate medication to complete hierarchical effect propagation. This design summarizes individual drug effects into collective effects aligned with patient conditions, which are then integrated to generate effective prescriptions. Moreover, PACRM adopts a two-stage optimization to sequentially learn candidate relevance modeling and structured effect summarization. Experiments on MIMIC-III and MIMIC-IV demonstrate that PACRM consistently outperforms competitive baselines, remains robust across varying sizes of medication sets, and is compatible with existing methods that leverage medication references.

Personalized electrocardiogram (ECG) generation is to simulate a patient's ECG digital twins tailored to specific conditions. It has the potential to transform traditional healthcare into a more accurate individualized paradigm, while preserving the key benefits of conventional population-level ECG synthesis. However, this promising task presents two fundamental challenges: extracting individual features without ground truth and injecting various types of conditions without confusing generative model. In this paper, we present \textbf{ECGTwin}, a two-stage framework designed to address these challenges. In the first stage, an \textit{Individual Base Extractor} trained via contrastive learning robustly captures personal features from a reference ECG. In the second stage, the extracted individual features, along with a target cardiac condition, are integrated into the diffusion-based generation process through our novel \textit{AdaX Condition Injector}, which injects these signals via two dedicated and specialized pathways. Both qualitative and quantitative experiments have demonstrated that our model can not only generate ECG signals of high fidelity and diversity by offering a fine-grained generation controllability, but also preserving individual-specific features. Furthermore, ECGTwin shows the potential to enhance ECG auto-diagnosis in downstream application, confirming the possibility of precise personalized healthcare solutions. Our code is available at \url{https://github.com/Raiiyf/ECGTwin}.

Transferring depth foundation models to endoscopic videos is promising for surgical geometric perception, but remains challenging under patient-level visual heterogeneity. Natural-scene depth models provide strong geometric priors, yet endoscopic imagery differs substantially in specular reflection, weak texture, tool occlusion, close-range anatomy, and acquisition conditions. Existing parameter-efficient adapters, including DV-LoRA-based methods, are typically static after training and apply the same adaptation to all frames, limiting their ability to handle input-dependent appearance shifts. We propose IC-DVLoRA, an input-conditioned dynamic adaptation framework for endoscopic depth foundation models. For each frame, a lightweight condition encoder extracts a compact token that captures appearance, texture, gradient, and structural cues. This token dynamically routes multiple DV-LoRA experts in a frozen transformer backbone and modulates a DPT-style decoder through feature-wise transformations, enabling sample-specific low-rank adaptation while preserving pretrained geometric knowledge. We train and validate IC-DVLoRA on SCARED and evaluate its zero-shot generalization on Hamlyn and C3VD without target-domain fine-tuning. Experiments show improved overall performance over static adaptation baselines and competitive results against recent endoscopic depth estimation methods. Further analyses of routing utilization, top-$k$ selection, trainable parameters, and inference latency demonstrate that IC-DVLoRA provides an effective and efficient adaptation strategy for endoscopic depth foundation models.

Real-world knowledge graphs evolve continuously, making retraining from scratch impractical. Continual knowledge graph embedding (KGE) updates representations with new triples while preserving historical knowledge. However, current methods face two challenges: (1) the stability-plasticity dilemma is pronounced for bridge entities connecting old and new subgraphs, where uniform updates may cause forgetting or rigidity; and (2) random negative sampling yields easy negatives and weak supervision near evolving graph boundaries. We propose KG-BRACE, a bridge-adaptive continual KGE framework combining topology-aware negative sampling with differentiated low-rank updates. KG-BRACE learns new entities via DoRA, updates bridge entities through a restricted Mini-DoRA branch, and freezes inactive historical entities. It also retrieves hard negative candidates from local neighborhoods. Experiments on six benchmarks show best or tied-best results on five datasets, consistent Hits@1 improvements, and competitive stronger freezing in extremely sparse graphs.

Few-shot knowledge graph completion aims to predict missing triplets involving few-shot relations, where each few-shot relation is associated with only a few examples of the relation, called support triplets. In few-shot knowledge graph completion, each few-shot task corresponds to predicting triplets for a few-shot relation. Most existing few-shot knowledge graph completion methods assume all entities and relations in knowledge graphs to be observed during training, making it challenging to deal with new entities and relations appearing at inference time. In this paper, we propose FIETA, a Few-shot InductivE knowledge graph completion method that introduces Task-Aware graph neural networks to handle few-shot tasks in knowledge graphs containing entities and relations unseen during training. Based on the local structures of support triplets, FIETA conducts task-centric modeling of relations to obtain relation representations reflecting relative positions to the few-shot relation, and performs task-guided message passing to compute entity representations optimized for each task by prioritizing triplets that are crucial in prediction. By learning to compute task-aware representations with structures, FIETA effectively handles few-shot relations on knowledge graphs with new entities and relations appearing at inference time. Experimental results show that FIETA outperforms 19 state-of-the-art baselines in various inductive scenarios.

\begin{abstract} Multi-modal Knowledge Graph Completion (MMKGC) leverages entity-associated visual and textual information to infer missing facts in multi-modal knowledge graphs. However, real-world multi-modal knowledge graphs often suffer from modality missingness. Existing methods typically rely on deterministic reconstruction or modality alignment, which fail to model the uncertainty and diversity of missing modalities and overlook the impact of intra-modality noise and structural semantics, thereby limiting their reasoning capability. In this paper, we propose Structure-Conditioned Missing Modality Generation (ScMMG), a generative framework for MMKGC that recovers missing entity modalities under knowledge graph semantic guidance. ScMMG leverages diffusion-based modality generation to model the latent distribution of missing modalities and incorporates adversarial structural optimization to enforce relational consistency within a unified semantic space. As a result, the generated modality representations are encouraged to be both distributionally plausible and structurally coherent. To provide reliable semantic anchors, we introduce a Fragment-level Magnitude-aware Fusion (FlMaF) module that suppresses noise and captures intrinsic entity semantics at a fine-grained level. Furthermore, a Structure-Conditioned Driven Diffusion Generation (ScDDG) module with Adversarial Optimization Scoring (AOS) is designed to enforce structural consistency and improve training stability within a shared semantic space constructed by RotatE, a rotation-based knowledge graph embedding model. Experimental results on DB15K, MKG-Y, and MKG-W demonstrate that ScMMG consistently outperforms state-of-the-art methods, while ablation studies confirm the effectiveness of each proposed component. \end{abstract}

Knowledge Graphs (KGs) are fundamental for grounding Large Language Models (LLMs) in factually consistent reasoning across diverse data-mining tasks. However, when reasoning over complex graph topologies, LLMs can be brittle in maintaining structural consistency and may rely on spurious correlations. We propose SKILL (Structural Knowledge Induced Logical Learning), a novel paradigm that transforms relational rules from passive symbolic patterns into explicit, generation-time symbolic guidance for LLM-based reasoning. Unlike existing methods that isolate symbolic mining from neural inference, SKILL introduces a semantically aligned rule synthesis framework that filters spurious patterns and enables LLMs to leverage relational regularities during inference. By grounding generation in pruned, query-specific subgraph structures together with rule–path evidence, SKILL improves inductive generalization while preserving generative flexibility. Extensive experiments on three inductive benchmarks demonstrate that SKILL achieves state-of-the-art performance, with up to a 5\% absolute improvement in Hits@1, highlighting a new and effective way to integrate symbolic structure into LLM reasoning over knowledge graphs. Source code is available at \url{https://github.com/luilui97/SKILL}.

Complex Query Answering (CQA) over incomplete Knowledge Graphs (KGs) formalizes compositional reasoning as Existential First-Order logic with one free variable (EFO\textsubscript{1}), where existential variables link multiple atomic constraints within a query. Maintaining existential consistency helps prevent reasoning shortcuts that satisfy only part of these constraints. We identify a Grounding--Skolemization dichotomy in existing CQA approaches. Grounding-based methods preserve logical equivalence but incur combinatorial enumeration costs, whereas Skolemization-based methods enable efficient latent reasoning over KGs while risking reasoning shortcuts arising from existential inconsistency. To address this tradeoff, we propose existentially consistent Skolemized reasoning, instantiated by the Logic-constrained Vector Symbolic Architecture (LVSA). LVSA uses vector-symbolic operations to compose atomic constraints and differentiable Skolemization to construct and reuse a constraint-conditioned latent state for each existential variable. Logical negation is handled through a logic-regularized neural negator. Theoretically, LVSA is structurally expressive for EFO\textsubscript{1} queries with acyclic query graphs and offers favorable complexity. Experiments show that LVSA outperforms prior Skolemization-based methods in reasoning performance and empirical existential consistency, while incurring substantially lower inference costs than Grounding-based approaches.

Numerical question answering (QA) over Knowledge Graphs (KGs) requires retrieving numerical values from structured sources. Knowledge Graph Question Answering (KGQA) and KG-augmented Large Language Model (LLM) methods exploit multi-hop connectivity for entity-centric questions but remain brittle for numerical queries, whose values typically appear as sparse attribute leaves. When required triples are missing, target values can become unreachable by explicit traversal; structural methods may retrieve values from the wrong numerical attribute, while numerical solvers assuming directly attached values cannot access non-local evidence. We propose \textbf{SEMAPHORE}, a neuro-symbolic framework that reformulates numerical QA over KGs as a \textbf{semantic detour} problem. SEMAPHORE combines symbolic path planning with embedding-based geometric navigation and decoupled scoring that separates relational navigation from numerical attribute grounding. This design identifies plausible detours while preserving the metric semantics required for accurate numerical answers. On numerical QA benchmarks derived from FB15K-237 and YAGO15K, SEMAPHORE outperforms strong LLM and KG-augmented baselines, demonstrating the value of semantic detours for robust numerical knowledge access over KGs. Sparsity and cross-dataset transfer experiments show that this advantage persists under severe graph incompleteness and when target-specific Knowledge Graph Embedding (KGE) training is replaced with transferred schema embeddings. Code is available at \url{https://github.com/orcies8147/SEMAPHORE}.

Knowledge graph question answering (KGQA) increasingly relies on graph-based retrieval-augmented generation (GraphRAG). However, graph retrievers are trained under path-level weak supervision, which treats every triple on an answer-reaching path as equally informative. Retrieved subgraphs thus contain spurious triples that reach the answer without causally supporting it, leaving noise filtering to the reasoner. We propose Causality-Guided Spurious Evidence Pruning with GFlowNet (CaseFlow), which internalizes spurious-evidence pruning into the retriever by reformulating retrieval as a flow-guided pruning process. Repurposed from a candidate generator into a pruner, the GFlowNet removes spurious evidence while retaining answer-supporting causal structure. To make this pruning causally informed and tractable, a relation-conditioned disentanglement encoder supplies a causal-relevance proxy for the chunk-level pruning prior, and adaptive relation-coherent chunks bound the action space. CaseFlow outperforms strong GraphRAG retrievers on the WebQSP, CWQ, and MetaQA-3hop benchmarks, with the largest gains in complex multi-hop reasoning. Our code is available at https://github.com/ChoiPi2/CaseFlow

Hybrid Question Answering (HQA) over semi-structured knowledge bases (SKBs) is vital for high-stakes domains like clinical precision medicine. However, HQA over SKBs is challenging because evidence is split across documents and relations, while answers must satisfy both semantic and structural constraints. Standard RAG pipelines retrieve evidence from textual or relational sources separately, failing to jointly capture the semantic conditions and relational constraints required by HQA over SKBs. Meanwhile, advanced agentic RAG systems that rely on computationally intensive models incur prohibitive inference latency and cost. To address these limitations, we introduce \textbf{RSA-RAG}, a \textbf{R}ole-\textbf{S}pecific \textbf{A}gentic \textbf{RAG} method with an asymmetric hybrid-path reasoning framework. We propose Role-Specific Asymmetric Collaboration for Hybrid-Path Reasoning, which assigns high-level roadmap planning and final answer generation to the large model, while lightweight role-specific agents handle frequent retrieval-control operations. Additionally, we introduce Strategy-Guided Candidate Navigation for Semi-Structured Retrieval, where compact agents perform strategy routing, constraint extraction, and hybrid navigation over an initial candidate set to retrieve textual and relational evidence around candidate targets. Furthermore, we implement Candidate-Support Evidence Structuring for Grounded Generation, which uses evidence verification to organize supported textual-relational evidence into structured grounding contexts for final generation. Extensive experiments show that RSA-RAG achieves strong retrieval performance on STARK, with an average MRR of 54.6\% and competitive top-ranked performance across the three subsets. It further improves CRAG accuracy by 11.1\% relative to Agentic RAG and reduces token consumption by up to 56.9\% compared with implemented LLM-intensive baselines. \end{abstract}

Bipartite graphs are a fundamental representation for relational data arising in recommendation systems, market-basket analysis, social networks, and communication graphs. A key challenge in these settings is to store and transmit large bipartite graphs compactly while preserving exact structural and path information. We study biclique-based representations of bipartite graphs $\boldsymbol{G}=(\boldsymbol{V},\boldsymbol{U},\boldsymbol{E})$, where the edge set is encoded using a collection of complete bipartite subgraphs. We focus on the Weighted Biclique Covering problem, which minimizes the total number of vertices used across all bicliques, and introduce a generalized variant that additionally penalizes the number of bicliques, capturing practical overheads in storage, transmission, and model complexity. While the weighted biclique covering problem is known to be $\mathsf{NP}$-Complete, we show that the generalized variant is also $\mathsf{NP}$-Complete. Despite this hardness, many real-world bipartite graphs admit low-dimensional geometric embeddings or can be well approximated by them. Leveraging this observation, we develop the first approximation algorithms with provable guarantees for the (generalized) weighted biclique covering problem on geometric bipartite graphs. Specifically, for $\delta$-disk bipartite graphs in low-dimensional $\ell_\infty^d$ spaces, we design a polynomial-time algorithm that achieves an $O(\log |\boldsymbol{U}| \cdot \log^d |\boldsymbol{V}|)$-approximation, combining ideas from greedy set cover, geometric range searching, and densest subgraph optimization. We also show how our algorithms extend to $\ell_\alpha^d$ metrics for any $\alpha\geq 1$. Finally, we evaluate our algorithms on real-world bipartite datasets and show that they efficiently compute significantly smaller biclique-based representations than natural baselines, while scaling to large graphs.

Graphs model complex relationships in domains such as social networks, bioinformatics, and the Web, where structures evolve continuously. Many studies focus on k-core community search for applications like recommendation and advertising. Due to frequent graph updates and the large scale of real-world graphs, data owners often outsource graph maintenance and execution to third parties. This raises concerns that third parties may forge query results intentionally or due to external attacks, making it essential to ensure the integrity of query results. Blockchain offers transparency and reliability for secure data storage, and hybrid-storage blockchains use off-chain storage with smart contracts for scalability. However, current authenticated query methods in hybrid-storage blockchains primarily focus on key-value data or other types of graph queries, which cannot be applied to authenticated community searches. This is the first study on the authenticated k-core community search in dynamic graphs. We propose a baseline scheme enabling clients to verify query integrity via verification objects (VO) and design a precomputation-based authenticated data structure, the Merkle Core Tree (MCTree), to enhance efficiency and reduce communication costs. Considering that both of the above solutions are based on Merkle hash trees (MHT), MHT-based updates require modifying paths from leaf to root upon each change, limiting update efficiency. To address this issue, we introduce HCoreGen, an HMAC-based algorithm that updates only affected tuples without precomputation, significantly improving update efficiency. The effectiveness and efficiency of authentication solutions are verified by the extensive experiment results on real large dynamic graphs.

Community discovery in heterogeneous networks commonly assumes that graph topology provides semantically useful neighborhoods. In human-centered systems, however, structural proximity may be weakly aligned with latent cognitive similarity, causing graph propagation to introduce psychologically inconsistent information. We revisit community discovery from this perspective and propose CogPHGNN, a cognition-aware graph learning framework that integrates multi-view psychological representations, semantic graph construction, and gated propagation. Extensive experiments demonstrate that graph learning becomes effective only when graph topology is sufficiently aligned with latent semantic and psychological similarity. The results show that graph construction quality and selective propagation critically affect cognition-aware community consistency, while mismatched neighborhoods or propagation mechanisms may degrade representation quality. Across graph constructions, propagation mechanisms, and psychological theories, the results consistently highlight structure-cognition alignment as a key factor governing cognition-aware community discovery.

Modularity is a standard objective for community detection, comparing observed intra-community connectivity with its null-model expectation. Modern node embeddings capture structural and semantic proximity beyond raw topology, but replacing adjacency with representation-induced affinity while retaining the classical degree-based null model makes the observed term geometry-dependent and the expected term geometry-agnostic. This creates a null-model mismatch: the objective may reward affinity already expected from representation proximity. This paper introduces RAM, a representation-aware modularity objective with a distance-calibrated null model. RAM preserves the observed-minus-expected principle by calibrating expected affinity with both degree heterogeneity and representation distance. We analyse its relation to classical modularity and develop divisive and multi-level greedy optimisers. Experiments on seven real-world datasets and synthetic scalability tests show that RAM improves over topology-only, direct embedding-clustering, and deep learning-based baselines, remains competitive under classical modularity while improving community quality on large graphs, and scales efficiently. Ablations confirm that both degree correction and distance calibration are essential.

\begin{abstract} Community Search (CS) aims to retrieve query-dependent cohesive subgraphs from networks, playing a crucial role in modern applications such as computing power networks and biological systems. Given the high cost of labeling, unsupervised CS has emerged as a scalable paradigm. However, existing approaches often suffer from restricted structural receptive fields and the limitation of static node representations, which fail to perceive the evolving community boundaries during expansion.To address these, this paper explores the potential of unsupervised CS by proposing the ComPass framework. We first design an incremental graph augmentation strategy to generate views with enriched community information. These views are then processed via a multi-scale structural tokenization mechanism, which transforms the graph into hierarchical sequences that bridge fine-grained topological associations with coarse-grained community contexts. By feeding these sequences into a Transformer-based encoder, ComPass effectively captures discriminative community representations without supervision. During online search, an iterative score refinement mechanism enables the co-evolution of community structure and node scoring, mitigating semantic drift.Extensive experiments illustrate the promising performance of ComPass as an exploratory step. Notably, ComPass achieves a competitive advantage over the strongest baseline, TransZero, with an average F1-score improvement of 15.23\%. This exploration suggests that integrating incremental augmentation, multi-scale tokenization, and adaptive search offers a potentially effective direction for advancing unsupervised community search. \end{abstract}

Structural entropy ($H_2$) is often reported beside normalized mutual information (NMI) and the adjusted Rand index (ARI), inviting the assumption that a lower structural objective implies better agreement with external labels. However, these metrics need not agree: $H_2$ measures random-walk compression, whereas NMI and ARI measure agreement with an external partition. For a fixed graph, we show that if every partition that differs substantially from a reference labeling has an $H_2$ value at least $\gamma$ greater than that labeling's $H_2$ value, then any partition whose $H_2$ value is at most $\varepsilon<\gamma$ above the reference labeling's $H_2$ value must also be close to the reference labeling. We measure this closeness using variation of information (VI), a standard distance between partitions, and derive a corresponding lower bound on NMI. These VI and NMI guarantees require a positive $H_2$ margin. Theoretically, we derive exact, checkable gap criteria for specified merge/split perturbations in balanced stochastic block models and for specified contact-domain boundary shifts. These criteria characterize when $H_2$ tracks external labels under the specified perturbations. Empirically, a sampled perturbation diagnostic assesses whether a graph--label pair exhibits a positive margin. Across 11 synthetic and eight real benchmarks, the pooled Spearman correlation between $H_2$ and NMI is $-0.83$ and $-0.60$, respectively. We identify four failure modes associated with departures from this trend: near-binary strong homophily, free-$K$ over-segmentation, elongated-community fragmentation, and hub mis-assignment. A multistart local search makes the diagnostic practical. A Hi-C case study illustrates the method on two chromosome-19 contact maps. Code, experiment configurations, and saved result summaries are available at https://github.com/SuuTTT/se-when-tracks-labels.

Correlation Clustering (CC) is a natural formulation of clustering in combinatorial optimization, which uses a graph representation of the input and does not require a pre-specified number of clusters. Given $n$ objects and a pairwise similarity function, the goal is to cluster the objects so that similar objects are put in the same cluster and dissimilar objects are put in different clusters. Despite its versatility, existing CC algorithms suffer from significant scalability issues and are inherently transductive: i.e., the algorithm must be executed from scratch for any new problem instance. In this work, we bridge this gap by leveraging Graph Neural Networks (GNNs) to solve Inductive Correlation Clustering, a novel generalization of the CC problem designed to handle unseen graph instances. By learning to exploit common structural patterns and node features during training, our framework generalizes to new graphs drawn from the same distribution with minimal computational overhead with respect to standard algorithms. We demonstrate the effectiveness and scalability of our approach through extensive experiments. Our framework not only excels in the inductive setting, e.g., lowering the inference time up to $5$ order of magnitude, while maintaining an approximation ratio within $~10\%$ of the best baseline solution, but also achieves competitive results on standard (transductive) CC benchmarks. Finally, we showcase a practical application of our framework as a learnable pooling mechanism for graph classification. Our results indicate that our method serves as an efficient pooling layer, enhancing the ability of GNNs to capture hierarchical structural information in networks.

An increasingly important requirement for clustering algorithms is that decisions and outcomes need to be clearly understandable for human users. We address this requirement for the Correlation Clustering (CC) problem, in case the objects to be clustered are accompanied by labels. In CC one seeks to partition a set of objects into clusters based on pairwise similarity and dissimilarity information. Our proposed model only considers clusters that can be explained by a label tree, which is essentially a decision tree that partitions the objects according to labels. The goal is to find an explainable partition of the data that minimizes the CC objective function. We formalize this problem from a theoretical viewpoint, and prove several hardness of approximation results, also when constraining the size of the tree. Our proposed greedy method finds locally optimal splits, and works recursively in a top-down approach. To efficiently compute the best split, we carefully maintain certain counters. An experimental evaluation demonstrates its effectiveness. The proposed algorithm processes graphs with over million edges in a few minutes, can find the ground truth in synthetic data, and is able to produce interpretable clusters in real world networks.

Recommender systems have become increasingly ubiquitous in daily life. While traditional recommendation approaches primarily rely on ID-based representations or item-side content features, they often fall short in capturing the underlying semantics aligned with user preferences (e.g., recommendation reasons for items), leading to a semantic-collaborative gap. Recently emerged LLM-based feature extraction approaches also face a key challenge: how to ensure that LLMs possess recommendation-aligned reasoning capabilities and can generate accurate, personalized reasons to mitigate the semantic-collaborative gap. To address these issues, we propose a novel Content Understanding from a Collaborative Perspective framework (CURec), which generates collaborative-aligned content features for more comprehensive recommendations. CURec first adapts the LLM with recommendation objectives through post-training, equipping it with instruction-following and chain-of-thought reasoning capabilities. Next, we use the LLM-generated features to train the recommender model. Then, we treat the user–item prediction scores produced by the trained recommender model as reward signals to guide RL-based refinement of the LLM, thereby improving the quality of generated reasons and pattern. The corrected reasons are directly integrated into the recommender model to enhance comprehensibility and recommendation performance. Extensive experiments on public benchmarks demonstrate the superiority of CURec over existing methods.

Large Language Models (LLMs) have recently emerged as powerful backbones for recommendation. To better elicit their capabilities, reasoning has been widely incorporated to help LLMs interpret rich textual signals and improve recommendation accuracy. However, explicitly generating intermediate reasoning traces often incurs substantial computational costs, which limits practical deployment in real-world recommender systems. To address this challenge, we propose \textbf{\textit{SelfDR}}, a \textbf{Self}-\textbf{D}istillation from \textbf{R}easoning framework for LLM-based Recommendation. SelfDR distills an LLM's own reasoning-enhanced predictions to produce recommendations directly, improving recommendation effectiveness while maintaining inference efficiency. All components in the framework are built on the same base LLM, without relying on any external models. Specifically, the teacher recommender is constructed by training a reasoner with downstream performance as the reward, enabling it to generate targeted rationales that are later incorporated into the teacher’s input. A student recommender for direct recommendation, with the same underlying model, then learns from the teacher through self-distillation with a dynamic weighting strategy. Extensive experiments on three public datasets validate the effectiveness, rationality, and efficiency of SelfDR. Codes are available at https://github.com/JiangDeccc/SelfDistillation.

Recommendation systems increasingly adopt a two-stage pipeline, where an ID-based retriever retrieves candidates and an LLM-based reranker refines their rankings. To improve retrieval quality, reranker-to-retriever distillation is commonly used to transfer the reranker’s knowledge to the retriever. For practical deployment, however, this pipeline must continually adapt to evolving interests and incoming interactions. A naive solution is to repeatedly update the LLM reranker and distill its latest knowledge, but this incurs prohibitive costs. Updating the retriever alone is cheaper, but its limited capacity makes adaptation from sparse data difficult. We propose SCoRD, a continual knowledge distillation framework for LLM-based reranking pipelines under a non-stationary data stream. SCoRD introduces a semantic reasoning assistant that distills the LLM’s ability to infer underlying user intents into reusable intent-level guidance. It selectively distills reranker knowledge to the retriever on low-confidence sequences, guides retriever-only updates without repeated LLM inference, and feeds retriever-derived representations and intent-drift signals back to the reranker. Experiments on real-world datasets show that SCoRD enables effective and efficient retriever-reranker co-adaptation.

LLM-based recommender systems increasingly adapt frozen backbones with parameter-efficient adapters as user-item interactions arrive. However, continual updates can overfit transient feedback, amplify noise, and degrade long-term preferences. We therefore propose \textbf{SCAU-Rec}, a drift-conditioned adapter-control framework that uses an evolving reference coreset to detect preference shifts and trigger updates only when warranted. When triggered, SCAU-Rec modulates risk-aware alignment between consecutive model states and projects adapter gradients away from an online low-rank stability subspace, reducing interference with retained preferences. We provide theoretical analyses of interference control, bounded predictive shift, and robustness to approximate stability subspaces. Across multiple benchmarks, SCAU-Rec improves recommendation utility, retention, update efficiency, and robustness over representative continual adaptation and LLM-based recommendation baselines.

Continual recommendation aims to capture evolving user interests from streaming data but struggles with sparsity. LLM enhancers mitigate this with semantic knowledge, but naive integration creates a new conflict. We identify this as the \textbf{Stability-Plasticity-Cognitivity (SPC) Trilemma}, where generalized LLM semantic priors (Cognitivity) conflict with retaining personalized historical preferences (Stability) and adapting to individual interest shifts (Plasticity). To address this, we propose \textbf{T}rilemma-\textbf{R}esponsive \textbf{A}daptive \textbf{C}ontinual \textbf{E}nhancement for \textbf{R}ecommendation (\textbf{TRACER}). TRACER synergistically combines three specialized modules, each targeting stability, plasticity, or cognitivity, while preventing any single lemma from dominating. This holistic design enables semantic knowledge to support history retention and adaptation to evolving interests without disrupting continual learning. Across five real-world datasets, TRACER effectively harmonizes the SPC trilemma and outperforms state-of-the-art baselines by up to 14.38\%. Our code is available at \url{https://github.com/woo-joo/TRACER_CIKM26}.

Graph Neural Networks (GNNs) have become a dominant paradigm in recommender systems by modeling collaborative signals over user--item interaction graphs. Recent LLM-enhanced recommenders seek to complement such structural signals with semantic knowledge from Large Language Models (LLMs). However, most existing methods implicitly follow a direct semantic fusion paradigm, where LLM-derived representations are treated as reliable semantic features and directly fused with collaborative representations. We argue that this paradigm is incomplete: LLM semantics are informative, but not directly trustworthy for recommendation. Through controlled diagnostic analyses, we identify two failure modes of raw LLM semantics: Unreliable Semantic Signals, where representations are sensitive to task-irrelevant textual noise, and Semantic--Behavioral Misalignment, where semantic proximity does not necessarily imply behavioral relevance. To address these issues, we propose PSCRec, a model-agnostic framework for preference-conditioned semantic calibration in LLM-enhanced graph recommendation. PSCRec progressively transforms noisy LLM-derived priors into behavior-consistent semantic evidence by structuring semantic profiles, grounding them in dynamic user preferences, filtering unreliable semantic associations, purifying residual semantic redundancy, and aligning calibrated semantic representations with collaborative signals. Extensive experiments on three benchmark datasets show that PSCRec consistently improves multiple graph-based backbones and outperforms representative LLM-enhanced recommendation methods. Further ablation, robustness, semantic--behavioral calibration, and efficiency analyses demonstrate the effectiveness and practical feasibility of the proposed calibration framework.

Multi-criteria ratings provide user feedback on multiple aspects of an item, offering richer preference signals than a single overall rating. Early work mainly focused on predicting criterion or overall ratings, whereas practical scenarios require ranking items according to their overall utility inferred from sparse and heterogeneous criterion-level evidence, which constitutes the core problem of multi-criteria recommendation (MCR). This ranking-oriented formulation of MCR gives rise to two key challenges: learning reliable criterion dependencies under sparse interactions, and bridging the mismatch between semantic preference signals and collaboratively learned representations. To address these issues, we propose S2R-LMG, a scoring-to-ranking framework that jointly learns from semantic and collaborative views. Specifically, we construct a user–item–criterion tripartite graph to model criterion dependencies in the collaborative view, while using large language models (LLMs) to transform sparse multi-criteria feedback into criterion-aware contextual preference priors for users and items. These priors are then projected into the ranking space and coupled with collaborative representations via alignment and retrieval-based distillation, so that semantic guidance directly contributes to preference ranking. Extensive experiments on three benchmark datasets show that S2R-LMG consistently outperforms state-of-the-art models on top-$N$ ranking, achieving an average improvement of 4.84% across all datasets and evaluation metrics over the strongest baseline. The implementation code is available at https://github.com/nunu1995/S2R-LMG.

Some LLM-based recommendation rerankers are evaluated under an oracle protocol that guarantees the ground-truth item is present in the scored set, either by injecting it into the candidate list or by scoring it against sampled negatives. We show that this protocol overestimates realistic performance by 92-95% in NDCG@10 across three primary Amazon datasets, so oracle results do not transfer to deployment. The root cause is a recall ceiling: realistic retrieval covers only 2-19% of relevant items at K=100 across eight datasets in three domains (Amazon products, MovieLens movies, MIND news), placing a deterministic upper bound on any closed-candidate reranker's top-k NDCG (Theorem 1, E[NDCG@k] <= Recall@|W_pi| under LOO evaluation, where W_pi is the reranker's own candidate window). Under realistic retrieval, every optimisation strategy we tested fails to produce a statistically significant improvement over the CF baseline on our primary Amazon datasets: prompt engineering, model scaling across a 168x parameter range, sequential models, three supervised neural rerankers (LambdaMART, RankNet, MLP) under a disjoint train/eval user split, a LoRA-fine-tuned LLaMA-3.2-3B reranker, and a hybrid CF+BM25+Dense retrieval with reciprocal-rank fusion. Adding text-aware retrieval raises recall on Beauty but does not lift end-to-end NDCG. Giving the LLM the upstream CF ranks and scores directly--the fairest setting we can construct--narrows but does not close the gap to CF, and does so by making the LLM adhere more closely to the CF order rather than by adding information; an LLM+CF fusion tuned on the evaluation users themselves recovers CF's performance and no more. We propose the Recall-Aware Evaluation Protocol (RAEP) as a diagnostic: classify the regime by retrieval recall, then evaluate reranking only where the ceiling permits differentiation. In the retrieval regimes we can measure (Recall@100 of 2-19%), the path to effective LLM recommendation runs through retrieval improvement rather than reranker sophistication; we do not claim this ordering holds at the higher recall, richer feature sets, and online feedback available to production systems.

Augmentation has become a central technique for improving deep forecasting models, but classification-style transformations tend to break the coherence between the look-back window and its continuous future target. We describe a simple procedure that unfolds the joint input-target sequence into overlapping sliding windows, randomly reorders a controlled fraction of them-prioritized by a lightweight variance criterion-and reconstructs the sequence by averaging across the overlaps, producing synthetic samples with controlled variation while limiting temporal distortion. The procedure is model-agnostic, introduces only three interpretable hyperparameters, and achieves strong improvements over a comprehensive set of competing augmentations across nine long-term forecasting benchmarks with five backbone families (TSMixer, DLinear, PatchTST, TiDE, LightTS) and four short-term traffic benchmarks with PatchTST. Component-wise ablations, hyperparameter sensitivity studies, distributional-alignment diagnostics, probabilistic forecasting evaluation, and a transfer experiment to univariate and multivariate time series classification clarify the contribution of each design choice.

Recent studies have explored the use of pretrained large language model (LLM) representations for long-term time series forecasting.While direct LLM-based forecasting methods adapt time series into LLM-compatible inputs, recent multimodal approaches further use LLM representations as auxiliary semantic representations and introduce interaction mechanisms between time series and language branches. However, existing methods mainly focus on aligning or combining these representations, leaving underexplored how LLM representations should be selectively utilized for different forecasting inputs. In particular, not all LLM layers are equally informative for time series forecasting, and different temporal regions may require different predictive cues from LLM representations. To address these challenges, we propose SeleRT (Selective Retrieval of LLM Representations for Long-term Time Series Forecasting), a multimodal forecasting framework designed to selectively utilize LLM representations for time series prediction. SeleRT consists of two key components: Statistics-Aware Dynamic Layer Fusion (SADLF), which dynamically generates layer weights from input time series statistics, and Patch-Wise Learnable Query Retrieval (PLQR), which enables patch-grounded queries to retrieve information from the fused LLM representation. We evaluate SeleRT on eight standard long-term forecasting benchmarks under full-shot and few-shot protocols. Experimental results show that SeleRT achieves state-of-the-art overall performance, with particularly strong gains in the few-shot setting. Ablation and visualization analyses further support the effectiveness of input-adaptive layer fusion and patch-wise selective retrieval.

Long-term time series forecasting has made significant progress by leveraging multi-scale information to capture hierarchical temporal patterns and model long-range dependencies. However, temporal downsampling in existing multi-scale methods inevitably smooths detailed temporal fluctuations, and this information loss is further aggravated by their emphasis on dominant trends across scales, resulting in insufficiently expressive representations. To address this, we propose a Multi-Scale Wavelet Mixing (MWMixer) model, which incorporates a Bidirectional Frequency-Bands Mixing strategy to recover lost temporal details across scales, enabling complementary cross-scale information interactions. Then, a Dynamic Scale-Adaptive Fusion module learns time-varying weights for each scale to fuse multi-scale forecasts into the final prediction, enhancing the flexibility of multi-scale aggregation. In addition, a cross-scale consistency loss aligns each coarse-scale prediction with the interval-averaged fine-scale outputs, while a multi-scale supervision loss enforces prediction accuracy at each scale, promoting consistent learning across scales. Extensive experiments on seven real-world datasets demonstrate that MWMixer achieves competitive performance in long-term forecasting.

General time series forecasting models pre-trained on extensive datasets have recently gained significant attention for their strong domain generalization capabilities. Nevertheless, existing time series pre-trained models confront two critical challenges. First, previous approaches are insufficient in capturing the intricate relationships among variables from a global perspective. Second, the temporal relationships in time series are explored using a fixed patch size from a single frequency perspective that discards important information hidden in the data. To address these challenges, we propose a frequency-enhanced pre-trained model (FAME) for multivariate time series forecasting, which captures the temporal and variable relationships separately from a frequency perspective. In the temporal relation learning module, we design a multi-expert architecture that automatically determines the size of patches for time series and process them using the customized models based on the determined patch size. In the variable relation learning module, we model the variable dependencies from a global perspective in the frequency domain and propose to employ a frequency characteristic prompt which enhance the extraction of relationships between more correlated variables. Finally, the outputs of the two modules are aggregated to generate the final prediction. We pre-train FAME on the Time-300B datasets and extensive experiments on public real datasets demonstrate that our framework achieves superior performance. Our code is published for further research.

To advance time series forecasting, a wide range of methods—from statistical techniques to deep learning architectures—have been proposed to improve prediction accuracy; however, most still follow a \textit{fast thinking} paradigm that maps historical patterns to future values with limited explicit reasoning or interpretable intermediate steps, undermining adaptability to evolving temporal dynamics. Although slow-thinking LLMs exhibit strong multi-step reasoning, prompt-only solutions are constrained by cost, privacy, and insufficient domain alignment, motivating the training of LLMs to acquire \textit{slow thinking} time-series reasoning skills. In this paper, we propose \Ours, a reinforcement fine-tuning framework designed to enhance the multi-step reasoning ability of LLMs for time series forecasting. Training starts with supervised fine-tuning for warm-up adaptation, followed by reinforcement learning that improves the model’s generalization ability via reward-driven reasoning optimization. Specifically, we introduce a fine-grained multi-objective reward to provide training signals tailored to time series forecasting, and GRIP (group-based relative importance for policy optimization) to leverage non-uniform sampling for encouraging the exploration of effective reasoning paths. Experiments show that \Ours achieves competitive forecasting performance across diverse datasets \footnote{Code is at \url{https://github.com/ustc-time-series/Time-R1}}.

Textual context can enhance time series forecasting when future dynamics are shaped by external events. In realistic forecasting pipelines, however, such context is typically retrieved rather than synchronously observed. Retrieved text is often sparse, temporally misaligned, and entangled with stale or non-predictive background information, making conventional multimodal fusion prone to negative transfer. We argue that the central challenge in multimodal forecasting is not to design increasingly sophisticated fusion mechanisms, but to determine when external textual context should revise a numerical forecast.To address this challenge, we propose Temporal Contextual Event Fusion (TCEF), an asymmetric selective revision framework that allows external textual context to influence numerical forecasts only when supported by relevant event evidence. TCEF first distills raw retrieved text into discrete timestamped events and estimates their horizon-specific relevance by jointly modeling event semantics, temporal recency, and the current state of the target series. Rather than treating text as an unconstrained forecasting signal, TCEF employs a prior-anchored revision branch to generate bounded residual corrections over a reliable numerical baseline. A reliability-guided trust gate then adaptively modulates textual influence across forecasting horizons, attenuating stale, weakly aligned, or uninformative signals while preserving the predictive structure of the temporal backbone.Experiments across nine Time-MMD domains and six temporal forecasting backbones show that TCEF consistently outperforms numeric-only forecasting and performs strongly against recent multimodal baselines. Further analysis shows that these gains are primarily driven by selective event-level revision rather than by simply incorporating more textual context.

While diffusion models have achieved impressive success in probabilistic time series forecasting, directly applying conventional noise injection may disrupt the continuity inherent in temporal data. Fundamentally, such additive Gaussian perturbations may sever the underlying dynamic continuity, generating intermediate states that are misaligned with the system's inherent evolution and making the recovery of valid sequential evolution exceptionally difficult. To resolve this misalignment, we introduce Koopman-Guided Diffusion (KoopDiff), a novel framework that restricts the generative process within an evolution-aligned subspace. Grounded in Koopman operator theory, our method lifts the diffusion dynamics into a coherent subspace where complex temporal dependencies are approximately linearized. Specifically, we establish this dynamically stable subspace by optimizing a Dynamical Subspace Alignment (DSA) objective, with analytical Koopman predictions serving as a Dynamical Prior to guide generative trajectories. Both theoretical analysis and a case study demonstrate that this formulation aligns the propagated noise with the dominant transition directions of the system, effectively maintaining dynamic continuity throughout the generative process. Extensive experiments on eight datasets show that KoopDiff achieves state-of-the-art forecasting performance while substantially enhancing inference efficiency. Our code is available at https://github.com/khwhk/KoopDiff.

Time series forecasting with variable missing is a pervasive yet challenging real-world problem. Missing values from sensor malfunctions or communication disruptions severely degrade the performance of conventional models designed for complete data, limiting their practical utility. To tackle the dual challenge of simultaneous imputation and forecasting, we propose EvoGNN, a novel Mutually Evolving Graph Neural Network, centered on a Node-Edge Co-Evolution mechanism that moves beyond merely capturing time-varying dynamics. Its core innovation lies in establishing a closed-loop, iterative refinement process where adjacency relations (edge) and node representations dynamically and mutually enhance each other. This is realized through two synergistic modules: an Edge-Driven State Refinement (ESR) module, which fuses static prior knowledge with a dynamically learned graph to update node states, and a Node-Driven Adjacency Evolution (NAE) module, which evolves the graph adjacency by reasoning over refined node increments and global temporal contexts. A Recurrent Equilibrium Gate (REG) is further introduced to regulate the evolution dynamics, ensuring stable and convergent learning. Extensive experiments demonstrate that EvoGNN consistently outperforms SOTA baselines across various missing scenarios. We theoretically demonstrate the convergence, effectiveness and efficiency of the co-evolutionary process, providing analytical insight into reliability and scalability.

Comments and summaries play a critical role in program comprehension. However, high-quality summaries are often missing in large codebases. Existing code summarization methods improve linguistic fluency but still fall short in capturing program execution semantics. Furthermore, source-level retrieval methods struggle to identify code examples that are syntactically different but functionally relevant. To address these limitations, we propose BRAD (Bytecode Retrieval-Augmented Deliberation), a novel retrieval-augmented deliberation model for code summarization that integrates bytecode-level semantics with multi-pass deliberation. Specifically, \toolname{} (1) encodes bytecode control-flow graphs (CFG) with a Graph Attention Network (GAT) to preserve control flow semantics, and (2) retrieves exemplar drafts in the bytecode text space to capture execution-oriented functional cues. These are then fused with textual encodings in a deliberative refinement process to generate the final summary. We evaluate BRAD on a public Java dataset, and the results show that compared with the strong baselines, BRAD increases BLEU-1 by 18.7\%, BLEU-2 by 16.2\%, BLEU-3 by 8.5\%, BLEU-4 by 2.8\%, ROUGE-L by 13.4\%, BERTScore-F1 by 6.3\%, and CIDEr by 10.0\%.

Retrieval-Augmented Generation (RAG) can undermine empathetic dialogue when retrieved content is tonally or emotionally misaligned with the user's expressed state: standard models forward the retrieved advice mechanically even when it is factually correct, eroding the supportive rapport the conversation calls for—a failure mode we term knowledge intrusion. We propose CARE (Conflict-Aware Reasoning for Empathy), which reframes retrieval-augmented generation as a decision to withhold: rather than treating retrieved documents as mandatory inputs to the generator, CARE introduces an explicit rejection action and learns when to take it. CARE cascades three mechanisms: Intent-Gated Retrieval, a probabilistic gate that suppresses retrieval when the user only needs emotional validation; Latent Critique, an explicit chain-of-thought step that decides whether to use or withhold the retrieved document on tonal and emotional grounds; and Conflict-Aware DPO, a preference objective trained on teacher-distilled robust/compliant response pairs to sharpen the rejection decision. At 50% injected noise, CARE improves Gating F1 over Standard RAG by up to 33.1 points, achieves a Context Rejection Score (CRS) above 70, and preserves generation quality. Our code is available at https://github.com/Zhizitan76/CARE-CIKM2026.

Graph RAG on corpus graphs enhances retrieval by leveraging intermediate node content as contextual clues to uncover unretrieved oracle nodes. However, existing methods suffer from two critical blind spots, namely semantically blind graph expansion and topology blind pruning, or else rely on prohibitively slow retrieval and generation interleaving. To address this, we formalize these limitations through an operational taxonomy and propose a retriever design that couples semantics-\textit{aware} adding with graph-\textit{aware} pruning under efficiency constraints. We instantiate this design as \textbf{GraFine}, which alternates between two refinement stages: Semantic Proximity eXpansion (SPX) for semantics-aware node addition, and a Graph Smoothing Reranker (GSR) for graph-aware pruning. Experiments on reference networks and text-rich knowledge graphs show that GraFine improves retrieval accuracy and generation quality while maintaining time-efficiency. Furthermore, we introduce Topological Recall~(TR), a metric that quantifies the topological proximity between retrieved and oracle nodes. Our analysis using TR confirms that GraFine more effectively steers refinement toward undiscovered oracle nodes by leveraging intermediate contextual clues. Our code is available at this link: \href{https://github.com/asmath472/GraFine}{\textcolor{navyblue}{\textbf{github.com/asmath472/GraFine}}}.

Multimodal retrieval-augmented generation (RAG) systems often rely on long unstructured contexts or aggressively expanded evidence graphs, which can introduce noisy evidence, weaken multi-hop reasoning, and increase unsupported generation. We present \textbf{GraphLoom}, a reliability-calibrated multimodal knowledge-graph RAG framework for compact and faithful evidence routing. Given a question and its associated multimodal input, GraphLoom constructs an instance-level multimodal knowledge graph from grounded scene descriptions, extracted relational triples, and external commonsense knowledge. Instead of injecting all retrieved evidence into the generator, GraphLoom performs reliability-aware subgraph retrieval with bounded expansion and selectively routes high-utility evidence through hierarchical graph memory slots and joint graph--sequence attention in a frozen language model. To improve robustness in complex reasoning settings, GraphLoom further combines interleaved retrieval with budgeted corrective retrieval, enabling adaptive multi-hop evidence refinement under noisy retrieval conditions. We evaluate GraphLoom on \textit{ScienceQA}, \textit{MultiModalQA}, and \textit{OK-VQA}, including large distractor evidence pools that approximate noisy external knowledge retrieval. Experimental results show consistent gains in answer quality and evidence faithfulness over strong multimodal RAG, graph-retrieval, and open-source vision--language baselines, with improved retrieval quality on MultiModalQA and stable performance under noisy evidence pools. Additional analyses using MiniCheck-based verification, human evaluation, and latency profiling show that reliability-calibrated graph evidence routing provides an effective alternative to long-context multimodal evidence injection.

Current work on improving reliability in large language model (LLM)- generated answers has primarily leveraged Retrieval-Augmented Generation (RAG), knowledge-graph augmentation, and reinforcement learning. While these methods are adept at enhancing and measuring reliability through semantic similarity and faithfulness, they often struggle to distinguish semantic similarity from geographic validity. This is especially critical in natural hazard management domains where geographic granularity (i.e., town vs. city vs. state) is significant for decision-making, as responses valid in one municipality may not transfer to another. In such domains, a confidently wrong answer carries greater risk than abstaining. We present GeoRisk-RAG, a novel hierarchy-aware framework that addresses this geographic-validity gap through selective answering. This framework explicitly estimates geographic applicability using a Directed Acyclic Graph (DAG)-based distance for context retrieval before response generation. Experiments on a novel held-out wildfire-related question-answering (QA) dataset show that GeoRisk-RAG significantly reduces false confidence rates for location-dependent questions, lowering the rate to 0.009 compared with ~0.090 for standard semantic similarity and reranking baselines, while consistently achieving higher human preference alignment. This work provides a more comprehensive assessment of end-to-end RAG pipelines by integrating geographic validity and selective-answering behavior for safer decision-making in geospatial domains.

While retrieval-augmented generation has become a common approach for enhancing question answering systems, retrieval is not universally advantageous. We study the problem of predicting whether incorporating external retrieved information is likely to improve response quality for a given question. To this end, we evaluate a range of prediction methods that are based on retrieval signals, answer characteristics, and semantic consistency between generated responses and retrieved passages. We further devise a predictor that probes the LLM's internal state. Its prediction performance significantly narrows the performance gap between post-generation methods which are computationally demanding and pre-generation (post-retrieval) methods. We use the prediction methods to devise a selective retrieval framework that dynamically chooses between retrieval and non-retrieval generation modes per question. Experimental results demonstrate that selectively applying retrieval augmentation yields answer quality that transcends that of using retrieval for all queries.

Retrieval-augmented generation (RAG) is typically optimized for topical relevance, yet its success ultimately depends on whether retrieved passages are useful for a large language model (LLM) to generate correct and complete answers. We argue that such utility is often LLM-specific rather than universal, due to differences in models’ knowledge, reasoning, and ability to leverage evidence. We formalize LLM-specific utility as the performance improvement of a target LLM when a passage is provided, compared to answering without evidence. To systematically study LLM-specific utility, we construct a benchmark of LLM-specific gold utilitarian passages for four LLMs (Qwen3-8B/14B/32B and Llama 3.1-8B) on six QA datasets (Natural Questions, TriviaQA, MS MARCO-FQA, HotpotQA, FEVER and 2WikiQA). Our analysis shows that utilitarian passages are model-dependent and non-transferable: each LLM performs best with its own utilitarian evidence, while evidence optimized for other LLMs is consistently suboptimal. Human-annotated evidence remains a strong general baseline but does not fully match individual LLM utility needs. We further introduce the LLM-specific utility judgment task and construct the corresponding benchmark, i.e., SpecUBench (LLM-Specific Utility Benchmark). Experiments show that existing utility-aware selection and scoring methods largely capture model-agnostic usefulness and struggle to reliably estimate LLM-specific utility. Overall, our findings highlight the limitations of current utility-aware retrieval and motivate generator-tailored evidence selection for improving RAG.

Multi-hop retrieval-augmented question answering depends on compact evidence sets whose members complement one another rather than repeat already covered facts. In evidence selection over a candidate pool, prevailing relevance-based selectors score each candidate mainly as a standalone match to the question, which obscures its marginal role under the evidence already selected. A candidate that appears weak in isolation may complete missing support after another unit establishes a bridge, while a highly relevant candidate may become redundant once its fact has been covered. This paper introduces Marginal Evidence Coalition Selection (MECS), a state-conditioned evidence selector that treats the partially selected evidence set as a coalition and estimates the marginal value of adding each remaining candidate. MECS derives pairwise preferences from public supporting-fact annotations over partial evidence states, and its typed scoring channels align candidate value with support addability, bridge continuity, and coalition consistency. This formulation permits candidate value to vary across evidence states, promoting candidates that complete uncovered support while suppressing repeated or incompatible evidence. Evaluated as an evidence selector over shared candidate pools, MECS improves support recovery and downstream answer accuracy on HotpotQA, 2WikiMultihopQA, and MuSiQue with frozen readers. Mechanism and robustness diagnostics further support the role of state-conditioned marginal selection in recovering complementary evidence for multi-hop QA.

Micro-view order dispatching (MICOD) is a critical task in ride-hailing platforms, aiming to optimize driver--order matching within limited spatiotemporal contexts. Prior works predominantly employ Reinforcement Learning (RL) to learn delayed matching policies. However, RL-based approaches often suffer from sample inefficiency and instability due to the need for extensive trial-and-error exploration. Furthermore, the combinatorial nature of dispatching typically breaks differentiability, preventing end-to-end alignment between prediction models and optimization objectives. In this paper, we propose DML, an End-to-End Delayed Matching Learning framework. Unlike RL, DML formulates order dispatching as a learning-from-expert problem inspired by imitation learning. By leveraging the insight that a global offline optimal matching (OPT) can be computed from historical data, DML constructs an expert in hindsight and avoids the need for online exploration. To effectively clone this expert behavior while enabling end-to-end training, we address two key challenges: expert decomposition and combinatorial differentiability. First, we construct a Delay-Augmented Bipartite Graph (DABG) and theoretically prove that the global OPT can be equivalently decomposed into local sequential learning targets. Second, we incorporate a differentiable optimization layer based on the Sinkhorn--Knopp (SK) algorithm and introduce a Batch-SK (B-SK) implementation to efficiently handle variable-sized graphs in parallel. Experiments on a real-world dataset collected from Didi Chuxing demonstrate that DML significantly outperforms state-of-the-art baselines in both matching quality and training efficiency.

In many data-mining applications, including recommender systems, influence maximization, and team formation, the goal is to pick a subset of elements (e.g., items, nodes in a network, experts to perform a task) to maximize a monotone submodular utility function while simultaneously minimizing a cost function. Classical formulations model this tradeoff via cardinality or knapsack constraints, or by combining utility and cost into a single weighted objective. However, such approaches require committing to a specific tradeoff in advance and return only a single solution, offering limited insight into the space of viable utility–cost tradeoffs. In this paper, we depart from the single-solution paradigm and examine the problem of computing representative sets of high-quality solutions that expose different tradeoffs between submodular utility and cost. For this, we introduce $(\alpha_1,\alpha_2)$-approximate Pareto frontiers that provably approximate the achievable tradeoffs between submodular utility and cost. Specifically, we formalize the \textsc{Pareto--}$\langle f,c \rangle$ problem and develop efficient algorithms for multiple instantiations arising from different combinations of submodular utility $f$ and cost functions $c$. % Since the $(\alpha_1,\alpha_2)$-approximate Pareto % frontier can often have a large number of solutions, We also provide an adaptive search algorithm that computes only a small subset of points that collectively summarize the entire Pareto frontier. Our results offer a principled and practical framework for understanding and exploiting utility--cost tradeoffs in submodular optimization. Experiments on datasets from diverse application domains demonstrate that our algorithms efficiently compute approximate Pareto frontiers in practice.

Dynamic heterogeneous task assignment is a fundamental problem in graph-based decision systems, where a scheduler must allocate continuously arriving tasks to multiple agents with diverse capabilities under coalition feasibility, deadlines, and resource constraints. Existing optimization-based methods often incur high replanning costs, while learning-based solvers usually struggle with weak task-agent coupling, sparse delayed rewards, and limited generalization under dynamic arrivals and coalition requirements. This paper presents BASAR, an end-to-end autoregressive framework for dynamic heterogeneous multi-agent task assignment on graphs. BASAR combines bidirectional task-agent cross-domain fusion, sparse-attentive state refinement, and feasible-action masked pointer decoding to generate online assignments over revealed task nodes and agent-specific fallback actions. To improve learning in collaborative scheduling, BASAR introduces an cooperative reward that evaluates the revealed task set and redistributes team utility through marginal service credits. Experiments on dynamic task assignment benchmarks show that BASAR achieves strong deadline-aware performance under fixed task workloads and high task throughput under fixed makespan execution, while reducing total planning time by up to 98.7\% compared with the optimization-based baseline. Our code is available at https://github.com/wuwenjianw/BASAR.

Real-time routing for customized shuttle buses presents a complex multi-agent control problem characterized by stochastic demands and strict topological constraints. However, efficient dispatching is frequently hindered by coarse-grained station representations, uncoordinated agent interactions, and the neglect of discrete operational constraints. To address these issues, HGA-PAC is presented as an event-driven framework where decisions are triggered upon vehicle arrival at stations. The framework operates through three synergistic components: a Heterogeneous Graph Attention encoder for fine-grained spatiotemporal dynamics, an attention-based communication module to mitigate resource competition, and a phaseaware decoder to ensure physical feasibility via dynamic masking. Experimental results demonstrate that our method significantly reduces operational costs and achieves a superior trade-off between service rate and travel efficiency compared to state-of-the-art baselines.

Short-session personalization and contextual recommendation often must improve later decisions from only a few selected-feedback observations. Standard online evaluation makes this improvement hard to interpret, because post-prefix reward may reflect both credit already stored from early actions and new reward information acquired while evaluation continues. We study this attribution problem through frozen-state credit storage, an adapt-then-freeze view in which rewards may update episode-local state during a short prefix, but later decisions can only read the state left at the freeze point. This boundary changes the design target: selected feedback is useful only if it is written before freezing into a state that remains readable afterward. Building on this insight, we introduce \textsc{ReBEP}, a bounded episodic-plasticity family organized around the action-local structure of selected feedback. \textsc{ReBEP} pairs a shared context scorer with resettable action-local memory, writes reward evidence only to the selected action slot, and reads the same slot after freezing. Its local-evidence update extends a scalar reward trace while preserving bounded action-local execution. Across latent-context tasks, supervised-to-bandit anchors, stress settings, and a short-session recommendation anchor, \textsc{ReBEP} stores prefix feedback more effectively as frozen decision credit than scalar traces, recurrent states, plasticity baselines, and frozen-prefix sequence summaries under the same write-read boundary. Memory interventions and component ablations show that the advantage depends on the frozen memory itself, action-slot assignment, bounded projection, and local-evidence writing rather than generic state capacity.

Recent progress in offline reinforcement learning (RL) has been driven by expressive generative actors such as diffusion and flow-matching policies, which capture multimodal behavior in offline datasets. However, these actors require multiple denoising or integration steps per action and thus incur substantial overhead at every decision in deployment. In this work, we revisit where capacity should be invested in an offline actor--critic method. Since the critic is used only during training and is discarded at deployment while the actor runs at every decision step, allocating capacity to the critic rather than the actor is more favorable for inference-time efficiency. However, scaling MLP critics in offline RL is known to introduce several distinct instabilities that have, in practice, kept critics shallow. We identify three distinct failure modes that arise when critics are deepened in offline RL---optimization, bootstrap-noise amplification, and value-range drift---and address each with a corresponding ingredient: a residual MLP backbone, n-step bootstrap targets, and a categorical cross-entropy loss. Combining these ingredients with a lightweight deterministic actor, we propose LAC (Light Actor, deep Critic). On OGBench, LAC matches the strongest diffusion- and flow-matching baselines while achieving up to 4X lower inference latency, comparable to one-step distilled policies without distillation. Its critic recipe also transfers across actor parametrizations.

Auto-bidding is a crucial task in real-time advertising markets, where policies must optimize long-horizon value under delivery constraints (e.g., budget and CPA). Existing methods for auto-bidding rely on compact numerical state representations: while they can implicitly capture delivery dynamics, they offer limited support for explicitly representing and controlling high-level intent, evolving feedback, and operator-style strategic guidance in real campaigns. Meanwhile, Large Language Models (LLMs) offer a powerful method for encoding semantic information, it remains unclear when LLMs help and how to integrate them without sacrificing numerical precision. Through systematic preliminary studies, we find that (1) LLM embeddings contain bidding-relevant cues yet cannot replace numerical features, and (2) gains emerge only with careful semantic--numeric integration rather than naive concatenation. Motivated by these findings, we propose \textit{SemBid}, a novel auto-bidding framework that injects LLM-encoded semantics into offline bidding trajectories at the token level. SemBid introduces three semantic inputs: \textit{Task}, \textit{History}, and \textit{Strategy}. It injects these semantics as tokens alongside numerical trajectory tokens and uses self-attention to integrate them, improving controllability and generalization across objectives. Across diverse scenarios and budget regimes, SemBid outperforms competitive baselines from offline RL and generative sequence modeling, with more consistent gains in overall performance, constraint satisfaction, and robustness. Our code is available at: \href{https://github.com/AlanYu04/SemBid-CIKM2026}{github.com/AlanYu04/SemBid-CIKM2026}.

Auto-bidding systems strive to maximize marketing value while maintaining high compliance with efficiency constraints, such as Target Cost-Per-Action (CPA). While Decision Transformers offer powerful sequence modeling capabilities, their application to this setting faces two challenges: 1) standard Return-to-Go conditioning causes state aliasing by ignoring the cost dimension, preventing precise resource pacing; and 2) standard regression constrains the policy to mimic historical averages, limiting its capacity to optimize performance near the high-efficiency boundary. To tackle these challenges, we propose PRO-Bid, a constraint-aware generative auto-bidding framework featuring systematic redesigns across data, architecture, and training via two synergistic mechanisms: 1) Constraint-Decoupled Pareto Representation (CDPR) separates global constraints into recursive cost and value contexts to restore resource perception, while reweighting trajectories based on the empirical Pareto frontier to prioritize high-efficiency data; and 2) Counterfactual Regret Optimization (CRO) employs a global predictor to evaluate alternative actions and identify promising local adjustments. By utilizing these high-utility outcomes as weighted regression targets, the model overcomes mean regression and approaches the empirical high-efficiency boundary. Extensive experiments on two public benchmarks and online A/B tests show that PRO-Bid achieves better constraint satisfaction and value acquisition than state-of-the-art baselines.

Existing industrial-scale navigation applications contend with massive road networks, typically employing two main categories of approaches for route planning. The first relies on precomputed road costs for optimal routing and heuristic algorithms for generating alternatives, while the second, generative methods, has recently gained significant attention. However, the former struggles with personalization and route diversity, while the latter fails to meet the efficiency requirements of large-scale real-time scenarios. To address these limitations, we propose GenMRP, a general framework for multi-route planning. In GenMRP, routes are generated iteratively. The first iteration identifies the optimal route, while the subsequent ones generate alternatives that balance quality and diversity using the newly proposed correctional boosting approach. Each iteration incorporates road features, user historical sequences, and previously generated routes into a Link Cost Model to update road costs, followed by route generation using the Dijkstra algorithm. Extensive experiments show that GenMRP achieves state-of-the-art performance with high efficiency in both offline and online environments. To facilitate further research, we have publicly released the training and evaluation dataset. GenMRP has been fully deployed on the Amap app for over a year, and it has brought a substantial improvement in user trajectory coverage, demonstrating its effectiveness and benefits.

Accurate estimated time of arrival (ETA) prediction is central to on-demand delivery platforms, yet it must operate under rapid distribution shifts, heterogeneous geographies, merchant partners, and consumers, as well as long-tailed operational edge cases. At marketplace scale, ETAs represent a promise, where even small errors can compound into substantial impacts on consumer experience and operational efficiency. We present DoorDash's next-generation ETA system, built on four contributions. First, a Feature-wise Linear Modulation (FiLM) based Base Model supporting point and probabilistic output heads. Second, a multi-task architecture that jointly produces consistent ETAs across surfaces and geographies. Third, we cast consumer-facing ETA window generation as a constrained optimization problem and decompose it into a two-layer architecture that separates accuracy-focused modeling from business-aware policy. Fourth, we pair probabilistic Base Models with an uncertainty-aware Decision Layer policy. The system is deployed in production at DoorDash's global marketplace scale and validated through large-scale online A/B experiments. Against our historical tree-based baseline, the deployed Base Model iterations reduce prediction mean absolute error by 34.6%, improve on-time accuracy by 19.9%, and improve ETA consistency by 37.6%, accompanied by material annual improvements in Gross Order Value. Layering the uncertainty-aware Decision Layer policy on top raises the error reduction to 35.0% and the on-time-accuracy gain to 20.9% in shadow deployment; online A/B validation of that policy is the next step in its rollout.

Graph anomaly detection (GAD) aims to identify nodes that deviate from normal structural or semantic patterns, with broad applications in financial transactions, network security, and online social platforms. However, existing GAD methods predominantly focus on sophisticated optimization within the graph domain, such as refining neighbor selection and aggregation strategies, and thus inevitably suffer from aggregation anomaly dilution. This issue is particularly pronounced under weak-information scenarios with sparse connectivity and ambiguous features. Moreover, such methods largely overlook the rich semantic cues in raw textual content. To unleash the textual modality for compensating insufficient graph signals and mitigating aggregation anomaly dilution, large language models (LLMs) have emerged as a promising solution given their strong semantic reasoning. Nevertheless, effectively fusing textual embeddings with graph structures remains challenging. In this paper, we propose LAGAD, the first framework that introduces the LLMs-as-Consultants paradigm into GAD. LAGAD leverages LLMs to extract external knowledge from textual information to compensate for the insufficiency of graph signals, thereby enhancing GAD performance. Specifically, LAGAD identifies low-confidence nodes via dual-phase selection and builds representative two-hop subgraphs via semantic similarity. The encoded subgraphs, together with node-level textual and structural descriptions, are fed into the LLM for structure–semantic joint prediction with natural-language explanations. The LLM predictions are then converted into pseudo-labels for local retraining, forming a closed feedback loop that refines the base model. Extensive experiments demonstrate that LAGAD effectively mitigates challenges under weak-information settings and significantly enhances the performance and robustness of underlying GNN models.

Industrial fraud detection often relies on costly expert-crafted features that overlook graph-structured relational signals, while GNNs often do not meet the interpretability and deployment requirements of financial risk control. We propose GraphFAS (Graph Feature Automated Selection), a distributed feature selection procedure based on Boruta that bridges this gap through: (1) a non-parametric graph feature generation module that constructs explicit, interpretable structural features via multi-hop subgraph extraction and multi-scale aggregation without learned parameters; and (2) an automated distributed feature selection algorithm extending Boruta with median-based aggregation across partitions to robustly identify informative features at scale with minimal domain expertise. Compared with end-to-end GNN pipelines, GraphFAS decouples feature aggregation from model training, enabling direct integration with tabular models and direct compatibility with TreeSHAP-based explanations. Deployed in Alipay, GraphFAS delivers order-of-magnitude improvements in engineering efficiency while showing strong performance against expert-driven and graph-learning baselines on large-scale graphs.

Voucher abuse poses a major challenge in e-commerce, where malicious users exploit promotional vouchers for profit. Unfortunately, fraud patterns evolve rapidly over time and across regions, causing distribution shifts that degrade existing detection models unless retrained frequently. To tackle this, we propose the Coupled Attribute–Topology Invariance Learning framework (CATeye). The key challenge arises from coupled attribute–topology shift, where edges built from attribute proximity cause environment-driven attribute shift to induce shifted topology, thereby amplifying variant signals through GNN message passing. CATeye sees through such coupled shifts with two learnable selectors. First, an Attribute Invariance Selector (AIS) learns node-adaptive masks to filter out non-invariant attributes. Then, conditioned on retained invariant attributes, an Edge Invariance Selector (EIS) samples an invariant subgraph and isolates non-invariant edges. Using the resulting invariant and non-invariant components, CATeye constructs multiple views and applies view-specific objectives to emphasize domain-invariant representations while suppressing domain-specific variations. Experiments on both a proprietary dataset from Lazada, a major Southeast Asian e-commerce platform, and a public benchmark show that CATeye consistently outperforms nine strong domain generalization and graph anomaly detection baselines, achieving up to an 8.61\% improvement in average F1 score over the strongest baseline. Source code is publicly available at https://github.com/Tian0426/CATeye.

Detecting merchant fraud is a critical challenge for global payment networks such as Mastercard, where data is inherently a hybrid of high-resolution tabular features and complex graph topology. While Graph Neural Networks (GNNs) based methods have shown promise in capturing the dependencies between nodes, deploying them at scale is challenging due to i) signal dilution under extreme heterophily, where extremely sparse fraudulent nodes are overwhelmed by benign neighborhoods, ii) prediction instability, where standard stochastic neighborhood sampling techniques for scalability leads to inconsistent risk estimates of merchants across inference runs undermining trustworthiness; and iii) high computational load at billion-edge scale. We present Surrogate-Guided Selection (SGS), a practical framework for stable and scalable fraud detection that combines strong tabular fraud models with graph neural networks. SGS introduces a dual-surrogate architecture, where lightweight gradient-boosted tree models are trained for merchants and cards to produce deterministic risk estimates. These strong dual surrogate scores jointly guide risk-aware deterministic neighborhood edge selection, and explicity accounts for node heterogeneity, yields task-aligned neighborhoods, produces consistent merchant representations, and naturally reduces the effective graph edge set. We evaluate SGS on Mastercard's production merchant-card network with 1.99 billion edges and 65 million nodes. SGS shows a 3.6% AUCPR lift over strong production baselines and consistently outperforms state-of-the-art fraud and heterophily-aware GNNs. Crucially, SGS achieves prediction stability while reducing total end-to-end running time by approximately 84%. We further show that SGS generalizes across two public graph anomaly benchmarks, improving performance without modifying underlying GNN backbone.

Online payment platforms generate massive transaction streams that contain rich behavioral signals for financial risk assessment. However, building a transaction foundation model for such tasks faces three major challenges. First, risk signals exhibit multi-scale temporal semantics, spanning immediate intents at the transaction level, short-term rhythms at the daily level, and long-term cycles at the monthly level. Second, financial transactions are inherently noisy and bursty. Third, an inherent objective gap exists between self-supervised reconstruction pre-training and downstream discriminative risk classification, often leading to inefficient adaptation and even negative transfer. In this paper, we propose a novel transaction foundation model, a pre-training and fine-tuning framework for large-scale financial risk assessment on transaction sequences. Specifically, we introduce a Hierarchical Transaction Transformer to explicitly model transaction sequences across transaction, daily, and monthly dimensions to comprehensively capture heterogeneous risk patterns at multiple granularities. To encourage the model to learn stable, essential behavioral representations, we design a Multi-scale Financial Summary Reconstruction Task that focuses on reconstructing aggregated statistical summaries rather than raw, noisy transaction values. To align the pre-trained representation space with downstream discriminative risk signals, we introduce an Intermediate Adaptation Stage that enables smooth knowledge transfer and significantly improves task adaptation. Extensive offline experiments and real-world online deployment on Tencent's large-scale payment platform demonstrate the effectiveness of our framework, providing a new paradigm for industrial transaction risk assessment.

Corporate default prediction is a core problem in financial risk management, yet traditional credit models rely heavily on financial statements that are often sparse or unavailable for many firms. Corporate transaction networks offer a complementary view of real economic activity, but how risk propagates through buyer–seller relationships remains underexplored. We conduct a large-scale empirical study using real-world electronic tax-invoice data spanning six years that links transaction histories with default events, revealing that transaction-driven risk is both role-dependent (buyer or seller) and scale-dependent. Based on these findings, we construct multiplex buyer-view and seller-view transaction networks and propose DefaultGNN, a dual-perspective graph neural network-based framework for corporate default prediction. DefaultGNN integrates both views to model how risk flows through transactional relationships, achieving strong improvements over both attribute-based and graph-based baselines, especially for firms with limited intrinsic risk signals. We further provide interpretable network-based explanations by visualizing how distressed trading partners contribute to default risk. In collaboration with a licensed credit rating agency, we validate that DefaultGNN's predictions complement existing credit scoring models, improving approval rates by 7-11%p without increasing default risk among approved firms. The source code can be found at https://github.com/jhkim611/DefaultGNN.

Tires are critical components that directly influence \textit{core vehicle dynamics}, and wet grip is a key performance criterion. However, real-world wet grip evaluation is \textit{significantly time-consuming}, and Finite Element Analysis (FEA)-based simulation still requires \textit{more than 10 hours} per design candidate, limiting rapid screening. To address this problem, we propose \textbf{\ours{}}, a \textbf{GR}aph-based \textbf{S}emi-supervised learning framework for \textbf{W}et gri\textbf{P} prediction designed for real-world development with scarce wet grip labels. \ours{} constructs a \textit{tire-graph} whose nodes represent tires, footprint images form node features, and edges connect tires with \textit{similar} design specifications. It then performs \textit{unsupervised graph contrastive pretraining} on all available tires, followed by \textit{supervised fine-tuning} on the limited labeled samples. Experiments on real-world tire development data show that \ours{} reduces prediction error by approximately 63.7\% in terms of RMSE relative to FEA-based simulation, while accelerating prediction from \textit{over 10 hours to 1.5 milliseconds} per sample. The source code is publicly available at \url{https://github.com/Bigdasgit/GRSWP}.

\begin{abstract} Machinery fault detection (MFD) remains heavily reliant on supervised learning, which struggles with the scarcity of fault labels in real-world settings. While reinforcement learning (RL) offers a framework to model the sequential nature of degradation, current ``RL-based'' MFD methods reduce the problem to a static contextual bandit (CB) formulation: by ignoring state transitions and discarding the temporal discount factor, they collapse to standard supervised classification. We propose an adversarial inverse reinforcement learning (AIRL) framework that treats MFD as an offline IRL problem. Unlike reconstruction-based approaches that rely on static error margins, or CBs that ignore dynamics, our method recovers an intrinsic ``health'' reward directly from observational state transitions, requiring neither manual reward engineering nor fault labels. On three run-to-failure benchmarks (HUMS2023, IMS, XJTU-SY), AIRL is the only method pairing within-window detection with non-saturated post-detection consistency across all datasets, while CB baselines fail to detect gradual degradation and reconstruction models collapse into always-anomalous states. Code and data: \href{https://github.com/dhirajneupane/AIRL-MFD-DN}{https://github.com/dhirajneupane/AIRL-MFD-DN}. \end{abstract}

Multiplayer Online Battle Arena (MOBA) games rely on matchmaking to maintain competitive balance. Our prior work, CUPID, framed matchmaking as an assignment re-optimization problem and showed that a single-mode win-rate predictor can meaningfully rebalance teams. However, deploying such a system across diverse player populations exposes three practical bottlenecks: most queueing players lack sufficient in-mode match history (cold start), skill distributions shift drastically across rank tiers (distribution inconsistency), and extreme skill segments are severely data-starved. We present CHAMP, a cross-domain matchmaking framework that resolves these deployment bottlenecks. To address data sparsity and cold starts, CHAMP replaces the target-mode-only player profile with a hybrid domain feature collection: a timestamp-ordered cross-mode short-term sequence whose slices are annotated with target-domain features, plus per-mode breakdowns of long-term, real-time and team statistics. We further propose the Domain-Aware Win-rate Network (DAWN): a Domain-aware Knowledge Extractor (DAKE) compiles target-mode attributes into learnable representations that feed Domain-Aware Temporal/Spatial/Permutation OmniNet Encoders (DATOE/DASOE/DAPOE), so that mode-conditioned representations and per-mode debiasing are learned jointly inside a single shared network. Online, one trained DAWN serves every supported mode, with per-mode position-satisfaction thresholds as the only mode-specific knob. Offline, DAWN achieves 67.73\% win-rate prediction accuracy, outperforming all evaluated attention and sequence baselines. Online A/B tests across the entire League ladder of a large-scale MOBA game, from novice players up to the top-expert players served by Elite Mode, demonstrate consistent drops in imbalanced matches. For lower-tier players, CHAMP reduces the 5-minute kill crushing rate by up to 20.73\%.

Accurate geological characterization of subsurface reservoirs from well log data is essential to support projects such as carbon capture and storage (CCS), geothermal development, and extraction of natural resources. Existing automated techniques for geological characterization primarily use sliding-window classification, which limits their ability to understand broader geological contexts, often leading to misaligned formation layers. To overcome these limitations, we introduce LithoFormer, a robust framework for stratigraphic inference using a Seq2Seq transformer model that ingests entire multivariate well logs in a single pass. It features a PatchTST backbone, rotary positional embeddings (RoPE), and a decoupled multi-task head for predicting geological zones and boundary probabilities simultaneously. To ensure robustness, we implement a constrained data augmentation strategy and a geology-informed loss function that enforces domain constraints. Validated and deployed on three real-world datasets, LithoFormer demonstrates a 90% reduction in median boundary error and eliminates stratigraphic order violations compared to traditional sliding-window baselines.

User tracking aims to identify network activities belonging to the same user across websites and network environments, supporting applications such as personalized recommendation and targeted advertising. As dynamic address allocation and modern privacy protections constrain traditional tracking methods, traffic-based tracking techniques have emerged to link dynamic addresses through traffic patterns. However, existing methods usually rely on large-scale user labels for supervised training, while such labels are difficult to obtain in TLS encrypted traffic, severely limiting model generalization. In this paper, we propose SLATE, a self-supervised address linkage framework integrated with LLMs under limited labeled data. Specifically, SLATE introduces a homologous subgraph discrimination task based on address behavior partitioning, which encourages the encoder to capture stable behavior patterns across temporal evolution from large-scale unlabeled data. Moreover, SLATE develops a graph semantic projection module, which enables the LLM to better interpret behavior graph representations through graph-text matching and node semantic classification, and further leverages its reasoning capability to support address linkage decisions. We evaluate SLATE on two real-world user datasets. Experimental results show that SLATE substantially outperforms the strongest self-supervised methods, achieving an average improvement of over 15\% in Hits@1, while remaining competitive with state-of-the-art supervised methods. These findings demonstrate the strong potential of combining self-supervised learning with LLMs for address linkage in tracking ecosystems.

Multi-party data collaboration increasingly requires deduplicating datasets held by independent providers. Although multi-party private set union (MPSU) provides a natural privacy-preserving abstraction, existing protocols require substantial provider-side interaction, while delegated MPSU still relies on a provider for result reconstruction and incurs super-linear computation overhead. We propose PrivDedup, a lightweight and fully outsourced protocol for exact privacy-preserving multi-party deduplication. PrivDedup recasts deduplication as protected table aggregation: providers independently encode encrypted records into invertible bloom lookup tables (IBLTs), and the server reconstructs the deduplicated union by aggregating and decoding the submitted tables without online inter-provider coordination. To prevent leakage beyond the intended union, pseudorandom coefficient scaling hides cross-provider overlap counts, while pairwise bucket masking conceals provider-specific table contents and provenance. If primary peeling leaves a residual core, a compact auxiliary table under an independent hash family enables recovery of the remaining elements. Evaluated over dataset sizes from $2^{10}$ to $2^{20}$ and 3--10 providers, PrivDedup reduces communication by 1.39--8.92$\times$ and end-to-end computation time by 1.13--12.24$\times$ compared with representative state-of-the-art MPSU and delegated MPSU protocols.

Cloud-hosted Large Language Models (LLMs) have become key knowledge sources for domain-specific model training. Prompt-driven Knowledge Distillation (KD) allows clients to leverage these models for student training, but client prompts often contain sensitive information that is exposed when sent in plaintext. Existing Differential Privacy (DP)-based methods protect teacher-side data rather than client prompts, while defenses that inject Gaussian noise into prompt embeddings remain vulnerable to reconstruction attacks and may degrade distillation utility. To address this gap, we propose \textsc{VeilKD}, a scalable framework for \emph{dual-private prompt-driven knowledge distillation} that protects client prompts via secure inference while preserving server-side model confidentiality. \textsc{VeilKD} employs linear-attention to enable cross-prompt batching and homomorphic optimizations for linear and nonlinear layers, achieving high-throughput encrypted teacher inference. Experiments across representative teacher-student distillation workloads demonstrate that \textsc{VeilKD} effectively prevents prompt reconstruction, preserves student utility within 0.10 average points of plaintext KD without output-side DP and DP-KD under matched DP settings, and accelerates per-thousand-prompt teacher-signal generation by 43.03--179.96$\times$ over prior secure inference methods.

Modern Information Retrieval (IR) systems, including small and medium-scale ones, often search across millions of documents. In many cases, data owners outsource retrieval to a third-party host to keep the data accessible. Furthermore, IR has also moved toward semantic matching with dense embeddings; although these embeddings are less interpretable than sparse representations, they remain vulnerable to model inversion attacks. Sending document embeddings to a third-party retrieval provider can, therefore, create privacy risks. An honest-but-curious host may try to reconstruct the original documents and exploit their content. One option is to send encrypted or obfuscated embeddings instead of the original vectors. Existing methods, however, have important limits: Asymmetric Scalar Product-preserving Encryption (ASPE) is vulnerable to inversion, while homomorphic encryption is often too costly. Building on the ASPE literature, we propose the Pivot-based Intersection Noise Scheme (PINS), a noise-injection method designed for dense IR. PINS aims to preserve retrieval quality while reducing leakage from document embeddings. It selects a set of pivots to define an obfuscation space, then chooses or samples an obfuscation point in that space for each document embedding. When measuring privacy-utility trade-off in terms of AUC, PINS reaches up to twice the effectiveness of state-of-the-art embedding obfuscation methods and also works with asymmetric bi-encoders, overcoming the current state-of-the-art. To favour reproducibility, we release PINS at \url{https://github.com/guglielmof/2026-CIKM-PINS}.

Retrieval-Augmented Generation (RAG) is a promising technique for applying LLMs to proprietary domains. However, retrieved documents may contain sensitive knowledge, posing risks of privacy leakage in generative results. Thus, effectively erasing private information from retrieved documents is a key challenge for RAG. Unlike traditional text anonymization, RAG should consider: (1) the inherent multi-document reasoning may face de-anonymization attacks; (2) private knowledge varies by scenarios, so users should be allowed to customize which information to erase; (3) preserving sufficient publicly available knowledge for generation tasks. This paper introduces the privacy erasure task for RAG and proposes Eraser4RAG, a private knowledge eraser that effectively removes user-defined private knowledge from documents while preserving sufficient public knowledge for generation. Specifically, we first construct a global knowledge graph to identify potential knowledge across documents, aiming to defend against de-anonymization attacks. Then we randomly split it into private and public sub-graphs, and fine-tune Flan-T5 to rewrite the retrieved documents, excluding private triples. Finally, the PPO algorithm optimizes the rewriting model to minimize private triples and maximize public triples retention. Experiments on four QA datasets demonstrate that Eraser4RAG achieves a strong privacy--utility balance among document-level rewriting methods, outperforming GPT-4o-based rewriting in private knowledge erasure while preserving substantially more public knowledge than aggressive keyword- or entity-suppression baselines.

The scaling of Large Language Models (LLMs) heavily relies on high-quality proprietary data, yet strict privacy regulations (e.g., GDPR) render vast amounts of real-world text containing Personally Identifiable Information (PII) inaccessible. Existing sanitization techniques struggle with implicit reference leakage and contextual degradation. In this paper, we propose a paradigm grounded in the hypothesis that LLM reasoning capabilities stem predominantly from the invariant logical context rather than from specific PII entities. We introduce a Graph-Driven Context-Preserving Anonymization Framework comprising three phases: (1) Global Graph Construction to explicitly model dependencies, (2) Targeted Topological Knock-out to isolate and prune explicit subject nodes and implicit coreference edges, and (3) Contextual Re-alignment for autoregressive reconstruction. Across the proprietary CCC-500 corpus and the public Text Anonymization Benchmark (TAB), evaluated over three random seeds against 18 heuristic, transformer-based, differentially private, and LLM-driven baselines, our method reaches $100.0 \pm 0.00$\% Deterministic Desensitization Rate on CCC-500 and $99.8 \pm 0.10$\% on TAB under parser-completeness conditions, while preserving downstream Top-1 Accuracy within $0.1$ percentage points of unsanitized text on identity-orthogonal multi-hop QA. An isolation ablation confirms that both the graph topology and the constrained regenerator contribute independently to the result. All improvements over the strongest generative baseline (GPT-4o Few-shot) are significant at $p<0.001$ on a paired Wilcoxon signed-rank test.

The defining question for approximate machine unlearning is not how strongly to perturb the model, but which signal in its representation to perturb. Existing methods perturb both at once: the sample-specific detail that identifies the forget set and the semantic structure that supports generalization. The result is a familiar trade-off, where forgetting and utility move in opposite directions, with no principled rest point. We separate the two signals through feature-map quantization: the quantized feature retains the semantic structure while shedding the identifying detail, and therefore specifies what an unlearned representation should look like for each forget sample. Quantized Contrastive Unlearning (QCU) uses the logit of this quantized feature as the positive anchor and the logit of the original feature as the negative in a contrastive objective, pulling the model toward a target that is, by construction, close to the retrained-from-scratch model. The closeness is not asymptotic. We provide an analysis showing that, under explicit small-drift and alignment assumptions, the QCU stage moves the model closer to the Retrain reference in expectation, yielding a tighter initialization-dependent convergence bound for subsequent retain-set fine-tuning. The same loss form, without modification, consistently closes the gap to the retrained reference on both utility and privacy across image classification (CIFAR-10/100, SVHN, ImageNet), LLM unlearning on TOFU, backdoor removal, LiRA-based privacy auditing, and prompt-level concept erasure in Stable Diffusion, while every baseline fails to close the gap along at least one axis. Code is available at https://github.com/hyunjoon0208/QCU.

As the Internet of Things (IoT) becomes deeply embedded in daily life, users are increasingly concerned about privacy leakage, especially from video, which holds rich, sensitive information. While many techniques can protect privacy, frame-by-frame protection in large-scale video analytics (e.g., smart communities) introduces significant latency. A more efficient solution is to first detect frames containing privacy objects (e.g., faces) and then apply selective protection. However, pixel-domain detection requires fully decoding video frames, incurring decoding overhead. Compressed-domain approaches directly analyze compressed video streams to accelerate inference, but often sacrifice accuracy, and the per-frame processing remains inefficient. Therefore, we propose ComPrivDet, an efficient framework for detecting privacy objects in compressed video by reusing I-frame inference results. ComPrivDet provides a pixel-domain detector on I-frames and a lightweight compressed-domain detector on P/B-frames. Leveraging compressed-domain cues to determine the presence of new objects, ComPrivDet either reuses I-frame results to skip P/B-frame detections or refines them with a lightweight detector. We prototype ComPrivDet on an edge server with an NVIDIA GeForce RTX 3090 GPU. ComPrivDet maintains $99.75\%$ accuracy for private face detection and $96.83\%$ for private license plate detection while skipping over $80\%$ of inferences. It averages $1.36\%$ higher accuracy with $57.23\%$ lower inference latency than pixel-domain baselines, and $10.93\%$ higher accuracy with $44.65\%$ lower latency than compressed-domain baselines.

Large-scale aerial imagery provides fine-grained visual information for urban analytics, traffic monitoring, emergency response, and geospatial data mining. A central challenge is to reliably detect tiny objects in high-resolution, dense, and cluttered scenes. Existing detector pyramids commonly rely on stride-based downsampling to aggregate multi-scale features. When objects occupy only a few pixels, such reductions may discard or alias fine spatial cues and make object representations sensitive to the sampling phase. We propose PolyphaseDet, a neck-local aggregation design for sampling-aware aerial tiny-object detection. PolyphaseDet contains two components. Polyphase-Preserving Injection (PPI) preserves phase-specific stride-4 evidence through space-to-depth reindexing before injecting it into the stride-8 representation. Phase-Offset Downsampling (POD) aggregates responses from multiple sampling offsets before committing features to a coarser grid. The backbone, decoder, matching strategy, loss function, and inference procedure are kept unchanged, isolating the effect of the proposed neck aggregation design. On VisDrone2019-DET, PolyphaseDet improves a controlled baseline from 28.8 to 32.9 AP at 640 × 640 input resolution, and reaches 40.3 AP at 1280 × 1280. On AI-TOD, it achieves 28.1 AP. Ablation studies show that the gains are not simply due to forwarding P2 features, adding extra branches, or using generic blur-based downsampling. PolyphaseDet also improves prediction stability under one-pixel input translations while maintaining real-time efficiency comparable to that of the RT-DETR baseline.

Remote sensing image fusion and semantic interpretation are usually optimized separately, so reconstruction quality need not preserve downstream retrieval utility. We propose LANCE, which uses structured language as an explicit, indexable semantic state. A frozen vision–language model maps the input to a closed-vocabulary text record; a language-conditioned generator fuses the imagery; and a dense verifier with a differentiable continuous summarizer re-describes the output. Text–text consistency and anti-collapse contrastive text–mask losses with negative queries close the loop without dense annotations or VLM fine-tuning. A variational EM formulation yields a per-step contraction result under regularity assumptions, distinguishing semantic faithfulness from verifier overfitting. Across seven test sets on GF2, QuickBird, WorldView-III, and WorldView-II, LANCE improves reduced-scale reference metrics by up to 11.7% (median 2.1%) over the strongest learning baseline. Independent evaluation using a held-out VLM and a supervised aerial detector yields gains of +6.4 mAP and +0.072 open-vocabulary F1 over the closest competitor. Text perturbations, cross-verifier transfer, and pseudo-RGB ablations examine the mechanisms underlying these gains. The framework connects image fusion to verifiable semantic representations for multimodal retrieval and knowledge management.

3D scene graph prediction is commonly optimized with object-node and relation-edge classification losses. These objectives encourage exact object and relation label recovery, but they do not directly optimize semantic proximity to the ground-truth scene graph: two wrong graphs can receive comparable penalties even when one is much closer in object roles and relation contexts. A direct graph-distance loss is difficult in general, since graph matching and graph edit distance require correspondence search and edit-cost design. Our key observation is that prior supervised 3D scene graph prediction protocols are conducted on fixed object-node indices and supervised edge slots. This can turn a scene graph comparison into a well-defined aligned graph-distance problem. To this end, we propose SWORD, a Scene-Grounded Wasserstein Object-Relation Distance that builds block-normalized object and relation ground distance metrics from training scene graph annotations through a deterministic base-to-context order, without external semantic priors. SWORD lifts these metrics to Wasserstein distances over aligned node and null-augmented edge marginals, providing semantic graph-distance supervision beyond exact-label penalties while preserving the ground-truth graph as the unique zero-distance solution in aligned marginal probability space. We prove that SWORD is a strict metric on aligned marginal probabilistic scene graphs, preserves scene-graph-grounded substitution ordering, controls additive graph-grounded semantic observables, and yields a surrogate upper-bounding decoded semantic error.

Generalized Category Discovery (GCD) aims to classify unlabeled data containing both known and novel classes, given only labels for known classes. Existing GCD methods typically assume a relatively balanced data distribution, an assumption that breaks down in realistic open-world scenarios. In this paper, we study Long-Tailed Generalized Category Discovery (LT-GCD), where both labeled and unlabeled data follow long-tailed distributions. Under such conditions, training signals are dominated by known categories; this imbalance biases representation learning and weakens novel-category discovery. To address these challenges, we propose VISRA, a semantic-driven framework for LT-GCD built upon a parametric GCD pipeline. Rather than relying solely on distribution debiasing, our framework leverages Vision-Language Models (VLMs) to provide additional semantic supervision for representative novel-class instances and guide visual representation learning. Experimental results on three long-tailed datasets demonstrate that VISRA achieves state-of-the-art performance, effectively alleviating novel-class supervision deficiency and learning more discriminative representations under severe imbalance. Our code and prompt are available at https://github.com/EileanZYZ/VISRA.

\begin{abstract} Recent advances in Large Language Models (LLMs) and vision--language models have introduced richer semantic guidance into Weakly Supervised Video Anomaly Detection (WS-VAD). However, video-dependent text generation or encoding during inference introduces substantial computational overhead and complicates deployment in resource-constrained scenarios. We propose APD-VAD, an Asymmetric Privileged Distillation framework that uses LVLM-generated descriptions as privileged information during training and deploys a visual student without dynamic text processing. A Privileged Language Cascade (PLC) supplies video- and snippet-level descriptions to the teacher, and a Background-Debiased Graph (BDG) attenuates scene-dominated relations before temporal graph reasoning. Semantic-distribution, feature, and topology distillation transfer the resulting class, snippet, and relational knowledge to the student. At inference, a dual-branch visual graph and Visual Prototype Memory (VPM) provide category-aware visual context through cached category prototypes and class-routed visual retrieval. Experiments on UCF-Crime and XD-Violence demonstrate competitive anomaly detection and improved fine-grained temporal localization, while deployment profiling confirms efficient inference without recurring LVLM processing. Code is available at \url{https://github.com/cgao-comp/APD-VAD}. \end{abstract}

Existing deepfake detection methods learn facial representations that capture the identity, semantics, or geometric structure of real faces. We argue that such face-centric representations can introduce identity and semantic biases, encoding appearance attributes that correlate with a person’s identity rather than with the image’s manipulation, thereby limiting transferability to unseen forgery types. We propose Forgery-Centric One-Class Artifact Learning (FOCAL), a representation learning framework that models manipulation artifacts directly rather than generic facial semantics. FOCAL is trained exclusively on forged images: an encoder-decoder reconstructs spatial artifact maps from fake inputs, forcing the encoder to learn where and how manipulation has occurred. This reconstruction objective is complemented by spatial and frequency-domain contrastive losses that encourage invariance to input perturbations while preserving discriminative forgery cues. Because the encoder captures a compact forgery-artifact distribution, real faces unseen during training naturally fall outside. We exploit this property by proposing the Dynamic-Centroid Mahalanobis Distance (DCMD), which enables classifier-free, zero-shot detection with out target-domain adaptation. FOCAL surpasses state-of-the-art methods on standard cross-dataset benchmarks in both detection AUC and pixel-level forgery despite fake-only training. This clearly demonstrates that forgery-centric representations yield transfer- able features than approaches anchored to real-face distributions

Deepfake manipulations evolve rapidly across digital media platforms, requiring continual detector updates as the distributions of face forgeries shift. Unlike general class-incremental learning, continual face forgery detection is a domain-incremental binary task in which new manipulations of the same fake class are introduced. We demonstrate that existing continual deepfake detectors remain tied to distribution-specific cues, preserving performance on seen datasets while struggling to generalize to unseen manipulation domains. This observation suggests that effective continual adaptation should preserve transferable representations while retaining prior knowledge under sequential domain shifts. We further find that SVD-based parameter-efficient tuning improves transferability but requires regularization against forgetting. Accordingly, we propose TASER, an exemplar-free framework that couples this transferable basis with transport-guided regularization. Asymmetric Class-Wise Partial Optimal Transport (AC-POT) aligns real and fake manifolds separately for class-aware adaptation, while OT-guided Contrastive Separation (OTCon) strengthens the real/fake boundary using transport-selected positives and opposite-class negatives. TASER achieves state-of-the-art intra-dataset retention and cross-dataset generalization, with a final average AUC of 0.9787, a 0.0123 average forgetting AUC, and an overall cross-dataset AUC of 0.8626 across challenging continual face forgery benchmarks.

Image-text emotion detection (ITED) aims to identify the emotion expressed by a given pair of text and image. However, existing methods often overlook color, a fundamental visual cue that carries rich emotional information. From the perspective of human cognition, color expressions in images and texts can evoke emotions in both direct and indirect ways. Inspired by this observation, we design a \textbf{D}ynamic \textbf{E}motion \textbf{S}emantic \textbf{E}nhancement \textbf{F}ramework via \textbf{D}ual \textbf{C}olor-induced \textbf{P}athways (\baby) for the ITED task. Specifically, our framework first builds a direct color-induced emotion (CIE) pathway based on general human color-emotion patterns and an indirect pathway through the generative imagination of color-similar scenes. Such rich CIE knowledge is then parameterized as multiple trainable experts within a mixture-of-CIE-experts module (\textsc{MoCE}) shared across the two modalities. However, as not all the color-induced emotions contribute positively to the final prediction, we design a dynamic expert selection mechanism that allows words in texts or regions in images to adaptively retrieve experts' knowledge consistent with the grounded emotions. Extensive experiments on three ITED datasets under different modality fusion strategies demonstrate the effectiveness, robustness, and generalization ability of \baby.

Multimodal Aspect-Based Sentiment Analysis (MABSA) aims to identify aspect-level sentiment from user-generated image-text posts. Although visual information can provide useful context for sentiment reasoning, it is not uniformly beneficial for aspect-level prediction. Different image regions may serve distinct roles: some provide content support for post understanding, some offer sentiment-related evidence, while others introduce irrelevant or misleading signals. To address this challenge, we propose VEGA, a MABSA framework that integrates visual evidence construction with evidence-guided region calibration. Specifically, VEGA first leverages a multimodal large language model to construct two complementary types of visual evidence from each image-text pair: content-aware visual evidence and context-aware sentiment evidence. These evidence streams are then used to calibrate region-level visual representations through a dual-support cross-attention mechanism, enabling image regions to contribute according to their roles in content understanding and sentiment inference. Finally, the calibrated visual features are integrated with the original text in a unified encoder-decoder architecture for both multimodal aspect sentiment classification and joint multimodal aspect-sentiment analysis subtasks. Experiments on public benchmarks demonstrate that VEGA consistently outperforms strong task-specific baselines. Further analyses confirm the effectiveness of visual evidence construction and evidence-guided region calibration.

\begin{abstract} Aspect Sentiment Triplet Extraction (ASTE) aims to extract aspect--opinion--sentiment triplets from review sentences and is a fundamental task in fine-grained sentiment analysis. Recent large language model-based chain-of-thought (CoT) methods improve instruction following and semantic reasoning for ASTE, but their verbose reasoning outputs introduce substantial inference overhead and often lack explicit structural grounding. We propose ReasonGraph, which integrates gold-guided CoT distillation with dependency-graph-based syntactic grounding. The distillation pipeline first trains the model with explicit reasoning chains and final answers, and then shifts to direct answer generation, allowing reasoning ability to be internalized while keeping inference compact. Meanwhile, a GATv2 dependency-graph encoder injects parsed syntactic structures as continuous prefix embeddings, enabling the LLM to adaptively balance syntactic and semantic cues. Experiments on four standard ASTE benchmarks show that ReasonGraph achieves strong performance with significantly faster inference than CoT-based reasoning methods. Further analysis indicates that our method effectively captures complementary semantic and syntactic information and focuses on the most informative cues for triplet extraction. Our code is available at \url{https://github.com/cscsysu/aste-cikm}. \end{abstract}

Dialogue-level Aspect-based Sentiment Quadruple Analysis (DiaASQ) aims to extract all (target, aspect, opinion, sentiment) quadruples from multi-turn multi-party dialogues, where sentiment elements may be scattered across different speakers and turns. Despite exploiting dialogue structures, existing methods face two fundamental bottlenecks: First, span representations relying on average pooling treat all tokens equally, which inherently dilutes critical boundary cues and introduces semantic ambiguity—especially when boundaries carry sentiment-bearing modifiers. Second, existing methods model dialogue structures at the coarse utterance level, neglecting fine-grained span-pair interactions across turns—crucial for linking scattered sentiment elements. To address these limitations, we propose: (1) Boundary-Aware Bilinear Attention Pooling (BAP), which replaces uniform averaging with boundary-guided attention, using bilinear projections to dynamically weight span tokens based on their relevance to start/end boundaries; (2) Role-aware Dialogue Householder Projection (RDHP), which refines span-pair representations via geometry-aware partial projections, projecting away dialogue-irrelevant directions while preserving sentiment-bearing subspaces. Extensive experiments on Chinese and English DiaASQ benchmarks demonstrate competitive performance, establishing a new state-of-the-art on the English benchmark with substantial gains in cross-turn quadruple extraction.

Named Entity Recognition and Relation Extraction are key Information Extraction tasks that enable the transformation of unstructured text into structured, machine-interpretable knowledge. In the biomedical domain, the extraction process must be guided by semantic constraints and domain knowledge to prevent the introduction of incorrect, inconsistent, or irrelevant information. In this paper, we propose a framework for constrained multi-stage information extraction pipelines guided by biomedical ontologies. The approach relies on the specification of a schema describing the entities, relations, constraints, and semantic dependencies relevant to the target domain. The schema is used to automatically generate a list of context-aware prompts that guide a general-purpose LLM during the extraction process (avoiding specific fine-tuning). The framework further supports mapping to ontology terms, extraction guidelines, and ontology-derived examples to improve extraction consistency and semantic validation. The proposed approach has been integrated into \schemaLink, an intelligent web platform supporting schema specification, prompt generation and execution, ontology-aware mapping, and visualization of extracted entities and relations. Extensive experiments on biomedical information extraction benchmarks show that our approach outperforms existing schema-guided prompt engineering methods by more effectively modeling dependencies between entity and relation extraction tasks, while achieving performance comparable to fine-tuned domain-specific models.

Decoder-only language models score text through next-token likelihoods over tokenizer-induced token sequences. This scoring interface becomes unstable when tokenizer segmentation shapes candidate selection or passage relevance scores rather than contextual fit alone. Agglutinative languages expose this problem sharply because morphosyntactic information is densely packed into surface forms. Korean offers a representative agglutinative stress case through particles, endings, compounds, and spacing variation. We propose Interleave Attention, a decoder augmentation that injects morphology-aware auxiliary representations into the scoring path. Prefix-aligned interleaving integrates these representations into a fused causal sequence rather than a separate cross-attention memory. Under matched initialization, continued-pretraining corpus, objective, and scoring interface, Interleave continued pretraining (CPT) produces stable trajectory-level gains over a parameter-matched decoder-only baseline, improving the overall benchmark average by 3.34 points and the context-conditioned average by 7.12 points. Controlled relevance scoring shows improved rank-sensitive scoring under a fixed hard-negative pool, and Korean surface-form stress tests indicate additional Korean-side robustness under surface variation. These results support a bottleneck-mitigation interpretation that separates the general Interleave effect from the morphology-based auxiliary-tokenizer effect.

Dynamic Embedded Topic Models (D-ETM) provide an interpretable framework for modeling temporal semantic evolution, but cross-corpus comparison remains difficult because topics are often learned independently and aligned only after training, a process that does not guarantee stable topic correspondence across corpora and time. To address this problem, we propose a D-ETM framework that first learns a common dynamic topic space over a merged multi-corpus collection, which we call the shared backbone, then introduces corpus-specific residual adaptation around the frozen backbone without creating separate latent topic spaces. This design preserves a shared topic index for cross-corpus comparison while allowing each corpus to specialize lexically. We evaluate the framework on three temporally structured corpora spanning 97 years: the Corpus of Historical American English, Harvard Business Review, and International Labour Review. Residual adaptation improves corpus-specific fit relative to the shared backbone while preserving the same-index cross-corpus topic trajectories, achieving substantially stronger alignment than full fine-tuning from the same backbone, with \(97.5 \pm 0.7\%\) versus \(17.9 \pm 1.1\% \) trajectory Retrieval@1, as well as stronger alignment than independent training with post-hoc Hungarian matching. These results suggest that incorporating topic alignment into the model can support more stable over-time cross-corpus comparisons while retaining corpus-specific lexical variation.

Generative recommendation leverages Large Language Models (LLMs) to model item semantics and user preference dynamics through sequence-to-token generation, but deep Transformer layers and wide Feed Forward Network (FFN) dimensions introduce substantial memory and latency overhead in online recommendation scenarios. Existing LLM pruning methods are mainly designed to preserve general language modeling ability, often overlooking the recommendation-specific states shaped by item semantics, collaborative signals, and temporal preference dynamics. To address this limitation, we propose Co-JDPRec, a collaborative joint dynamic pruning framework that shifts the pruning objective from parameter removal to recommendation-state preservation. Specifically, Co-JDPRec introduces a task-aware saliency criterion that combines collaborative alignment and sequence-order sensitivity, enabling the model to identify neurons critical for user-item relations and evolving preference dynamics. Guided by this saliency, Co-JDPRec jointly searches FFN width and layer depth to derive a compact heterogeneous architecture, followed by distillation recovery to restore recommendation performance. Theoretically, Co-JDPRec characterizes structural importance via first-order Taylor sensitivity under recommendation-aware objectives. Empirically, it achieves superior accuracy–efficiency trade-offs on both title-based and semantic-ID-based generative recommendation backbones, validating the effectiveness of recommendation-state-aware pruning.

Large Language Models (LLMs) have recently been explored for generative recommendation, where item tokenization plays a critical role. Existing studies design item identifiers such as discrete Semantic IDs, single-token identifiers, and multi-token or set identifiers. However, in many identifier-based methods, the item-level grounding of these introduced identifiers is not established during general-domain pre-training and must instead be learned during recommendation-specific adaptation. Consequently, LLMs may treat them as opaque prediction symbols rather than stable item representations. To address this issue, we propose MISTRec, a masked item-slot modeling framework for learning structured continuous item representations. MISTRec introduces three key designs. First, MISTRec decomposes each item into continuous CF-aware and latent semantic slots, reducing the entanglement between collaborative signals and metadata-derived item semantics. Second, it trains these slots with Masked Slot Autoencoding (MSA) by reconstructing corrupted historical slots and completing a fully masked future item. Third, it introduces False Negative Reweighting (FNR), which downweights semantically conflicting negatives during target alignment and improves the stability of full-catalog ranking. Experiments on three public benchmarks show that MISTRec consistently improves next-item recommendation over strong sequential and generative baselines, while yielding more stable item neighborhoods across random seeds.

In generative recommender systems, items are typically tokenized into fixed-length semantic ID sequences for autoregressive next-item prediction. However, for user-context modeling, this fine-grained representation triggers \textbf{Intra-item Attention Overload}: excessive attention is spent on low-level intra-item dependencies rather than high-level inter-item behavioral transitions. To address this, we propose \textbf{Semantic Subword Tokenization (SST)}, which represents historical items as variable-length semantic subwords while preserving fixed-length target decoding. SST first applies \textbf{Item-level Subword Tokenization (IST)} to merge stable adjacent \textbf{atom tokens} into compact \textbf{semantic subword tokens}, thereby reducing intra-item reassembly in the encoder. It then introduces \textbf{Behavior-induced Co-occurrence Augmentation (BCA)} to inject coarse-grained semantic prefix transition signals, guiding the freed modeling capacity toward inter-item behavioral regularities. Extensive experiments on three public datasets and three generative recommender backbones show empirical improvements of SST over fixed-length and transferable variable-length SID baselines. Code is available at \url{https://github.com/mxrcandy/Semantic-Subword-Tokenization}.

Recent generative recommendation models recast next-item prediction as the generation of a semantic identifier sequence. While this formulation enables autoregressive models to produce item IDs directly, the commonly used token-level likelihood objective does not distinguish between tokens that play different roles in item identification. This limitation is especially pronounced for semantic-ID representations, where prefix tokens often determine coarse item groups and later tokens provide finer-grained disambiguation. To better align training with the structure of semantic IDs, we study token-level learning signals from two complementary perspectives. First, we introduce a prefix-aware weighting scheme, Front-Greater Weighting, which emphasizes tokens according to their contribution to reducing semantic ambiguity among candidate items. Second, frequency weighting increases the learning emphasis on infrequent tokens, addressing the long-tailed distributions and popularity bias commonly observed in recommendation data. We further introduce a multi-target optimization framework with curriculum learning, which integrates the two token-weighted objectives with the standard likelihood and enables stable optimization with adaptive emphasis across training stages. Experiments on multiple benchmark datasets demonstrate that the proposed approach consistently improves generative recommendation performance over strong baselines and prior token-weighting methods. Additional analyses show that the method is robust across different semantic-ID constructions and backbone scales, and that it improves recommendation quality for both popular and long-tail items. Code is available at github repository.

Generative Recommendation (GR) has become a promising paradigm for large-scale recommendation systems, yet existing GR models typically perform single-pass decoding without explicit refinement, causing early deviations to accumulate and degrade recommendation quality. Recent work mitigates this via two reasoning paradigms: \emph{reason-then-generate}, which refines intermediate states before producing the draft, and \emph{reason-while-generating}, which interleaves reasoning with token generation. Both, however, reason before a complete draft exists and cannot diagnose or revise a finalized Semantic-ID (SID) trajectory. To this end, we propose GRC, to our knowledge the first structured reflection--correction paradigm for GR that performs reasoning on a complete draft via a \textbf{G}eneration–\textbf{R}eflection–\textbf{C}orrection process. GRC introduces a supervised reflection-correction template with multi-granular reflection labels in the SID space, and further optimizes the GRC trajectory via GRPO-based reinforcement learning with tailored rewards that jointly evaluate correction quality and reflection accuracy. For efficient online serving, we propose an Entropy-Guided Reflection Scheduling (EGRS) strategy that prioritizes high-uncertainty trajectories for correction under a fixed beam budget. Extensive experiments show that GRC consistently outperforms eight state-of-the-art baselines by up to 13.33\%, and online A/B tests demonstrate its substantial practical value in large-scale industrial recommendation, delivering a 1.39\% lift in advertising revenue with only modest latency overhead.

Predicting continuous values such as watch-time and gross merchandise value (GMV) is a core problem in industrial recommendation systems. Its inherent difficulty stems from the highly complex and long-tailed distributions of the target signals, which are hard to model accurately. Existing regression methods typically rely on fixed parametric assumptions on the target distribution: overly simple assumptions underfit real-world data, whereas more intricate ones tend to sacrifice scalability and generalization. To address these limitations, we propose a sequence modeling framework based on residual quantization (RQ), in which the target continuous value is decomposed into a sequence of quantization codes that represent progressively finer approximations. The model autoregressively predicts these codes from coarse to fine granularity, with each step refining the residual error left by the previous one. To further improve the quality of the learned representations, we introduce an ordinal-aware representation learning objective that aligns the RQ code embedding space with the ordinal structure of target values, thereby yielding continuous representations of quantization codes and more accurate predictions. We conduct comprehensive experiments on public benchmarks for watch-time and lifetime value (LTV) prediction, together with a large-scale online A/B test for GMV prediction on an industrial short-video recommendation platform. Across all settings, the proposed method shows competitive performance among existing state-of-the-art approaches and generalizes well across diverse continuous value prediction scenarios.

Self-supervised pre-training has emerged as an effective solution to data sparsity in Click-Through Rate (CTR) prediction. However, existing paradigms still face two critical limitations: (1) task degeneration in pre-training, where heuristic or random target construction may lead to trivial shortcuts; and (2) limited adaptivity in fine-tuning, where static transfer strategies fail to accommodate optimization heterogeneity across diverse feature fields and ID frequencies. To address these issues, we propose the \textbf{F}eature-aware \textbf{A}daptive pre-training and fine-tuning \textbf{F}ramework (FAF). The core of FAF lies in two coupled designs. First, we improve pre-training target construction from both token and feature levels. At the token level, we introduce a position-decayed future item prediction task to incorporate additional future behavior supervision while reducing the influence of noisy distant targets. At the feature level, we formulate masked feature selection as a combinatorial optimization problem and derive an entropy-redundancy scoring criterion, which selects informative target fields with low contextual redundancy to construct non-trivial reconstruction tasks. Second, we design an adaptive fine-tuning strategy that combines field-aware and item frequency-aware update regulation. This strategy dynamically adjusts embedding update magnitudes across heterogeneous feature fields and item frequencies, improving the compatibility between pre-trained representations and downstream task adaptation. Extensive experiments on three real-world datasets show that FAF brings consistent overall performance gains over strong baselines across multiple backbone models, demonstrating its effectiveness in representation learning and compatible knowledge transfer.

Graph Neural Networks (GNNs) have been adopted for click-through rate (CTR) prediction to explicitly capture and model feature interactions. However, existing graph evolution approaches with separate structural inference and feature aggregation face two key challenges in recommendation scenarios. On the one hand, recursively updated structural representations can suffer from degraded edge discriminability during graph evolution, destabilizing high-order topology learning. On the other hand, heterogeneous feature fields make conventional message passing prone to semantic misalignment. Consequently, the learned topology may not align with the semantic requirements of subsequent message propagation. In this paper, we propose Field-aware Compositional GNN (FC-GNN), which exploits field semantics as a dual-role inductive bias to stabilize graph learning. At the structure level, we introduce a Field Prior Matrix as a field-conditioned interaction operator, reducing the dependence of edge discrimination on feature representations. At the node level, we establish Field Semantic Anchors as shared semantic coordinates and employ a field-conditioned gating mechanism to calibrate heterogeneous feature representations before message passing. Moreover, we reuse the same field prior matrix during aggregation to form a feedback bridge between structure inference and semantic propagation. Experimental results on three public datasets demonstrate that FC-GNN consistently outperforms existing methods.

Negative sampling plays a pivotal role in implicit feedback-based collaborative filtering (CF) models optimized with pairwise ranking loss, as these models must learn to rank preferred items higher than non-preferred ones. As simply treating all unobserved interactions as negative samples may degrade model accuracy, previous studies have explored various negative sampling strategies, such as static and hard negative sampling, to select more informative negative samples. However, static negative sampling methods often fail to select informative samples because they do not account for training dynamics or the structure of user preferences, whereas hard negative sampling methods suffer from the false negative problem and may therefore introduce incorrect training signals. To address these limitations and provide a theoretical solution to the negative sampling problem, we conduct an in-depth analysis of implicit CF under the Bayesian Personalized Ranking loss. As a result, we find that sampling highly preferred items as negatives degrades recommendation performance, whereas more frequently sampling moderately preferred items improves it. Based on these findings, we propose Semi-Hard Negative Sampling (SHNS), a novel negative sampling method for implicit CF that selects moderately difficult negative samples while dynamically annealing the sampling difficulty to mitigate the risk of false negatives. Extensive experimental results show that SHNS achieves superior performance compared with 10 state-of-the-art negative sampling methods. Further empirical analysis demonstrates that SHNS effectively mitigates the false negative problem while maintaining the ability to select informative negative samples.

Negative sampling is a pivotal technique in implicit collaborative filtering (CF), enabling efficient and effective training by contrasting observed interactions with sampled unobserved ones. Recently, large language models (LLMs) have shown promise in recommender systems, yet their use in negative sampling remains underexplored. Existing methods often rely on textual information and task-specific fine-tuning, which may limit their applicability in settings with scarce textual metadata or constrained computational resources. To this end, we propose a text-free and fine-tuning-free \textbf{D}ual-\textbf{T}ree \textbf{L}LM-enhanced \textbf{N}egative \textbf{S}ampling method (DTL-NS). It consists of two modules: (i) an offline false negative identification module that leverages hierarchical index trees to transform collaborative structural and latent semantic information into structured item-ID encodings for LLM inference, enabling accurate identification of false negatives; and (ii) a multi-view hard negative sampling module that combines user–item preference scores with item–item hierarchical similarities from these encodings to mine high-quality negatives, thus improving models’ discriminative ability. Extensive experiments demonstrate that DTL-NS consistently improves recommendation performance across different implicit CF models, negative sampling methods, and LLMs.

KG-aware recommendation has been widely studied to alleviate data sparsity by using knowledge graphs (KGs), which represent items, entities, and their relations as graphs and provide item-side knowledge. However, existing methods incorporate item knowledge without considering how much each user or item node should rely on it. As a result, they apply KG signals indiscriminately across nodes, even to nodes whose collaborative filtering (CF) signals from the interaction graph (IG) are already reliable. In this paper, we propose AdaKG (Adaptive Node-Aware KG Fusion), a novel KG-aware recommendation method that adaptively adjusts the contribution of auxiliary knowledge for each node. Since user-item interactions and item knowledge provide different types of signals, directly mixing them can distort the CF signals. To avoid this, AdaKG separately encodes the IG and KG with view-specific encoders, allowing each view to capture its own information. It then estimates how strongly each node should rely on item knowledge by measuring the stability of its CF signals under small adversarial perturbations, assigning a larger KG contribution to less stable nodes. Finally, AdaKG adaptively aligns the IG and KG embeddings in a shared space and fuses them according to the estimated node-wise reliance. Through experiments, we show that AdaKG achieves strong performance compared with its baselines and the effectiveness of our adaptive fusion strategy.

User-user social networks are widely utilized as side information in recommendation and have been shown to improve performance. We argue that current social recommenders still have room for improvement. First, for social data utilization, most methods focus mainly on positive trust relations, while failing to effectively use negative distrust relations, which can also provide useful signals for recommendation. Second, in terms of modeling architecture, most methods rely on linear or nonlinear transformation structures (e.g., multi-layer perceptrons) as their foundational building blocks. Kolmogorov-Arnold Networks (KANs), recently introduced for their strong data-fitting ability, offer a promising alternative. To this end, in this paper, we propose SocialKAN, a novel KAN-based trust-distrust-aware social recommender method. Specifically, SocialKAN introduces three KAN-based graph learning channels to model the user-item interaction graph, user trust graph, and user distrust graph, respectively. To better enable effective cross-view adaptation among the three channels, we design a semantic routing module for knowledge sharing at the embedding level and an adaptive consensus knowledge distillation strategy for output alignment at the optimization level. Experimental results on the three datasets show that our proposed SocialKAN method achieves strong performance. Ablation studies further validate the rationality and effectiveness of the components designed in our SocialKAN.

Dense multimodal retrieval scales well, but it can miss relational evidence from user histories, item neighborhoods, and domain entities. We present KG-ARMM, a retrieval-compatible Residual Structural Fusion (RSF) framework. A text-and-poster backbone supports approximate nearest-neighbor (ANN) candidate generation; an offline item graph and filtered knowledge graph (KG) supply evidence to a lightweight pairwise re-ranker. The primary configuration gives the learned score limited authority, combining it with normalized structural support, a popularity penalty, and a long-tail bonus. On a temporal MovieLens–The Movie Database (TMDB) benchmark evaluated against the full catalog, the balanced variant improves normalized discounted cumulative gain at 20 (NDCG@20) by 8.3% (0.04273 to 0.04628) and Recall@20 by 7.0% (0.09968 to 0.10666) over the multimodal backbone. A matched full-scale control without graph/KG features or the direct structural term reaches 0.04405 NDCG@20 and 0.10336 Recall@20; complete-model paired gains remain positive with 95% intervals [0.00205, 0.00241] and [0.00285, 0.00375], respectively. KGSupport@20 rises from 0.2704 in that control to 0.3128, with stable catalog behavior. Sparse-history users gain most, and validation-holdout ablations favor the filtered relation bundle. Unrestricted learned-score authority yields higher accuracy but collapses long-tail exposure. These results support structural evidence as a sparse, auditable correction layer over dense multimodal retrieval rather than as an unconstrained replacement for the base ranking signal.

Multimodal recommendation improves item representations by combining visual, textual, and collaborative signals, but stronger cross-view alignment does not always improve ranking. Our diagnostics on Amazon Baby show that direct consistency has an effective range: moderate alignment helps, while stronger alignment suppresses recommendation-specific variation. We also observe a clear spectral split: low-frequency components capture shared structure, whereas higher-frequency components retain more private ranking signal. Based on these findings, we propose BRIDGE, a behavior-guided residual integration framework built on dual-frequency graph evidence. BRIDGE separates the model into three parts: DFGE decomposes graph-smoothed ID, visual, and textual channels into spectral bands; BEN converts training-only co-user overlap into signed candidate evidence; and CRI applies that evidence only inside the base top-$K$ candidate set during training and inference. This design keeps the multimodal backbone and localizes behavior evidence to candidate calibration. Experiments on Amazon Baby, Sports, and Electronics show that BRIDGE reaches 0.1128/0.1262/0.0778 Recall@20 and 0.0525/0.0594/0.0385 NDCG@20, outperforming baselines by up to 7.3\% in Recall@20 and 14.9\% in NDCG@20. Project materials are available at \url{https://lizesheng13.github.io/bridge/}.

Multimodal recommendation (MMRec) aims to enhance recommendation performance by leveraging rich item content from multiple modalities. However, directly incorporating all modality information does not necessarily lead to better preference modeling, since user interests are often more related to a subset of modality signals, while other signals may be weakly aligned with user preferences or even introduce noise. Although recent MMRec methods improve modality utilization through invariant learning, attention mechanisms, graph refinement, or contrastive learning, their alignment processes are often implicit or heuristic and lack a clear objective for selecting modality signals that better match user interests. In this paper, we propose AMUR, an information-guided selective modality-interest alignment framework for multimodal recommendation. Inspired by an information-theoretic view, AMUR aims to enhance modality information that is more related to user interests while reducing the influence of less aligned signals. Specifically, AMUR first refines modality graph structures towards user behavior, and then selectively aligns shared interest-related semantics across modalities. This enables AMUR to improve modality-interest alignment while preserving useful modality-specific complementary information. Extensive experiments on three real-world datasets demonstrate the effectiveness of AMUR over competitive baselines. \textit{\href{https://github.com/Wenze1/AMUR}{The codes are available here.}}

In Reinforcement Learning (RL)-based sequential recommendation, offline training often reduces user feedback to sparse binary rewards, weakening credit assignment and limiting policy optimisation. In visual recommendation settings, user choices are also shaped by product appearance, visual attention, and semantic impressions, which provide preference cues that binary outcomes cannot capture. However, how to transform such visual cues into reward supervision and how they impact recommendation performance remains underexplored. Therefore, we introduce ViSalRec, a unified visual-feedback reward augmentation framework for offline RL-based recommendation. ViSalRec treats visual preference cues as an additional reward layer that transforms item image, predicted saliency maps, and multimodal Large Language Model (LLM) judgements into dense reward signals, instantiated as saliency-based rewards, visually-conditioned LLM likeability rewards, and history-conditioned LLM purchase rewards. Within SA2C-style actor–critic recommenders, we study how these reward formulations interact with data characteristics, sequence encoders, and advantage-estimation strategies. Experiments on Amazon Review variants show that visual reward augmentation can improve recommendation performance, but its effectiveness depends on the proxy signal, data density, and reward-integration strategy. Overall, LLM-mediated and history-aware rewards are more reliable than raw saliency rewards, especially under denser data and value-based training. These findings provide an account of when visual preference feedback improves offline RL-based recommendation and how such rewards can be designed and integrated.

One major challenge of sensor-based Human Activity Recognition (HAR) is data heterogeneity, where domain shifts between training and testing data could significantly degrade recognition performance. Recently, large language models (LLMs) have demonstrated strong cross-domain generalization capabilities by modeling language with next-token prediction, offering a promising paradigm for mitigating domain shifts in HAR. In this paper, we propose NextHAR, a novel self-supervised sensor pretraining framework for generalizable HAR. NextHAR first tokenizes sensor data in the frequency space, converting them into motion-semantic tokens. Then, to learn human motion dynamics, NextHAR jointly models inter-channel correlations and temporal dependencies by predicting the next sensor token groups. Furthermore, we introduce a comprehensive cross-domain HAR benchmark to enable systematic evaluation. Extensive experiments on this benchmark show that NextHAR outperforms state-of-the-art methods by 13.83% in accuracy and 15.00% in F1-score. Moreover, NextHAR achieves an average inference latency of 0.033 seconds on wearable devices, highlighting its efficiency for real-world applications. Our code and organized benchmark are publicly available at https://github.com/saiketa/NextHAR.

Existing financial time series generation (FTSG) approaches predominantly model unconditional distributions, lacking the semantic controllability required to simulate deliberate market dynamics, such as macroeconomic shifts and geopolitical shocks. To bridge this gap, this paper studies text-conditioned FTSG that synthesizes realistic multivariate price sequences guided by natural language market descriptions. We propose \textbf{Fin-CoDiT}, a novel diffusion transformer framework engineered around a profound architectural inductive bias: the structural isomorphism between the hierarchical semantic depth of large language models (LLMs) and the spectral frequencies of time-series signals. To exploit this, Fin-CoDiT introduces a \textbf{Semantics-Frequency Alignment} mechanism via \textit{Grouped Routing Attention}, which dynamically routes explicit LLM layers to their targeted frequency bands during the denoising process. To encode cross-variable dependencies of time series without violating the temporal arrow, we propose \textbf{Time-Frequency Causal Variational Autoencoder}, which strictly operates as a channel-wise bottleneck equipped with orthogonal causal convolutions to fundamentally eliminate future information leakage. Furthermore, it incorporates a branched decoder with in-network differentiable morphological projections, guaranteeing absolute physical market constraints (e.g., $\text{Price}_{high} \ge \text{Price}_{low}$). For the training and evaluation of this novel paradigm, we construct an automated multi-agent workflow that creates high-quality text-time series pairs from real-world market states spanning January 1, 2015, to December 31, 2025. Extensive experiments on treasury futures and ETFs demonstrate that Fin-CoDiT establishes a new state-of-the-art, significantly outperforming existing baselines in both generation fidelity and controllability.

Many time-series generation studies train a separate generator for each dataset. We study a complementary setting in which one shared, channel-independent generator is trained on several timestamp-derived scenario pools and a compact textual prompt selects the desired pool at inference. We propose UPLOTS, a \textbf{U}nified, \textbf{P}rompt-guided \textbf{L}anguage model framework f\textbf{O}r constrained \textbf{T}ime-\textbf{S}eries generation. UPLOTS combines a Diffusion-TS generator with a truncated pretrained language-model backbone, a Time-series Prompt Embedding Module, and two multi-dataset balancing mechanisms: Curriculum-Aware Loss Modulation and Rolling-Loss Dynamic Sampling. We evaluate four real-world datasets under morning- and evening-peak scenarios and extend the analysis to 14 prompt configurations on ETTh and Energy, covering temporal, calendar, load-level, and volatility subsets. The results support unified, in-distribution generation for the predefined datasets and scenario prompts. Downstream forecasting experiments further assess the utility of the generated samples for data augmentation. Our code is available at \url{https://github.com/cruiseresearchgroup/UPLOTS}.

Missing data imputation in multivariate time series (MTS) becomes particularly challenging when an entire target node is unobserved throughout an input window, a setting we refer to as \emph{node blackout}. This setting arises in both spatiotemporal sensing datasets and more general MTS benchmarks, where a node denotes the basic relational unit and may correspond to either a physical sensor stream or a feature/variable channel. Existing methods increasingly exploit inter-variable dependencies through graph structures or attention-based interactions, but most still depend primarily on flat node-level support. Such support becomes fragile when blackout targets have weak local relations or when multiple blackout nodes occur within the same local region, making immediate neighborhoods insufficient or ambiguous for reconstruction. To address this issue, we propose \mname{}, a \textbf{H}ybrid \textbf{H}ierarchical \textbf{G}raph learning framework for MTS \textbf{I}mputation under node blackout. \mname{} preserves a sparse local graph as a stable structural prior while introducing a higher-level support graph over subgraphs. The higher-level support graph is constructed from two complementary views: inherited topology, which preserves topology-consistent support from the local graph, and subgraph similarity, which retrieves support from structurally compatible local subsystems beyond immediate node-level neighborhoods. To make this higher-level support adaptive to each input window, we construct it from temporal representations and stabilize subgraph formation with shared node priors. We further introduce a missingness-aware adaptive fusion mechanism that balances local and hierarchical supports according to node missing ratio and local-support reliability. Experiments on five real-world MTS datasets demonstrate strong and consistent performance of \mname{} across both explicit spatial settings and general multivariate benchmarks under node blackout. Our code is available at \url{https://github.com/jhBae00/H2GI}.

Time series anomaly detection aims to identify time points or intervals that deviate from normal temporal behavior. In reconstruction-based detection, this requires errors that distinguish anomalies rather than merely accurate reconstruction. However, most existing methods rely solely on temporal modeling, while vision-augmented approaches focus primarily on forecasting, leaving the potential of structural visual cues for reconstruction-based detection underexplored. Moreover, bridging temporal and visual representations further remains challenging. To address these challenges, we propose a Vision-Augmented Coordination framework for time series anomaly detection, called ViCO. A decoupled visual teacher and temporal student characterize normal behavior from complementary views: tri-view visual representations capture structural regularities, whereas raw sequences model value-sensitive dynamics. Prototype-Guided Coordination retrieves normal structural prototypes and conditions the temporal decoder, translating visual structure into reconstruction guidance for anomaly scoring. Extensive experiments on five real-world datasets demonstrate that ViCO consistently outperforms strong baselines by up to 9.0%, while enabling teacher-free inference for lightweight deployment. Code is available at https://github.com/Cedarrr1/ViCO.

Detecting emerging anomalous trajectories of unseen types is a critical yet challenging problem for applications such as urban planning, traffic management, and public safety. Existing methods often rely on closed-set assumptions, i.e., they either assume all anomaly types are known or detect anomalous trajectories without using anomaly labels, limiting their ability to detect novel, unseen anomalies. In this paper, we propose OpenTAD, an open-set trajectory anomaly detection framework for identifying anomalous trajectories of both seen and unseen anomaly types. A pseudo-anomaly generation strategy is introduced to address label scarcity by enriching the anomaly manifold, enabling OpenTAD to learn compact normal boundaries even with limited anomaly samples. To tackle spatio-temporal complexity, a memory-augmented trajectory encoding module is designed to capture high-order spatio-temporal dependencies under complex dynamics. Moreover, OpenTAD leverages distribution shift learning to handle anomaly uncertainty by maximizing the separation between normal and anomalous distributions in the latent space, ensuring reliable detection of unseen anomalies. Extensive experiments on multiple real-world trajectory datasets with both synthetic and real anomalies demonstrate that OpenTAD outperforms state-of-the-art baselines, achieving an $8.88\%$ improvement in detecting novel anomalies.

Online anomalous sub-trajectory detection is a core technology for applications such as real-time traffic monitoring and fraud prevention. Existing methods primarily suffer from three key limitations: 1) weak capability in fine-grained anomaly localization, 2) insufficient constraints regarding road network topology, and 3) difficulties in interpreting the spatiotemporal physical logic of anomalous behaviors. To address these issues, we propose OSTAR, an online generative framework for anomalous sub-trajectory detection driven by topology-aware autoregressive reasoning. Specifically, OSTAR first utilizes a macro-micro-aware perception module to capture local movement dynamics. Building upon this, it embeds road network topology into its autoregressive modeling to capture physical movement constraints. During inference, rather than outputting an uninterpretable anomaly score, it dynamically constructs a physically coherent reference path to serve as a transparent baseline. Finally, by decoupling the detection into independent spatial and temporal dimensions, OSTAR explicitly pinpoints "spatial detours" and "temporal abnormal dwelling." This design effectively overcomes the opacity of traditional black-box scores. Extensive experiments on two real-world datasets demonstrate that OSTAR achieves excellent accuracy in fine-grained anomaly localization while maintaining competitive inference efficiency, meeting the real-time requirements of industrial deployment. The source code is publicly available at https://github.com/lzxcompute/OSTAR.

Combining event and tracking data is fundamental to modern soccer analytics, yet the two sources are rarely well aligned: event timestamps recorded by human annotators often miss the true moment of the action, distorting the spatiotemporal context that downstream models rely on. Existing synchronization methods depend on noisy human-annotated event locations and fail to detect ball receptions, obscuring when each player gains ball possession. To address these limitations, we propose ELASTIC (\textbf{E}vent-\textbf{L}ocation-\textbf{A}gno\textbf{STIC} synchronizer), a framework that infers the start and end timestamps of events solely from player and ball trajectories, without relying on annotated event locations. To recover ball receptions, ELASTIC enriches the event sequence by inserting virtual termination events between consecutive events, so that the end of each event is detected jointly with its start. It then extracts a sparse set of candidate frames where ball touches are physically plausible, and aligns the termination-inserted event sequence with the candidate-frame sequence using an extended Needleman-Wunsch algorithm. For reproducible evaluation, we construct a publicly available benchmark by annotating ground-truth timestamps on the Sportec Open DFL Dataset, on which ELASTIC substantially outperforms existing methods. Through downstream task evaluation, we further show that improved synchronization translates into measurable gains in soccer analytics. The source code and benchmark are available at \url{https://github.com/hyunsungkim-ds/elastic.git}.

Integrating multi-source knowledge into urban knowledge graphs (UrbanKGs) has shown strong potential for improving urban spatiotemporal prediction (USTP). However, existing public UrbanKGs are often misaligned with the temporal and multi-scale nature of USTP. First, they usually represent urban facts as static triples, ignoring when facts become valid, expire, or change over time. Second, their functional-zone layers are typically derived from static POI clustering, which overlooks spatial boundaries, activity interactions, and event-driven temporal semantics. As a result, current datasets provide limited support for modeling temporally valid knowledge and its cross-scale effects on prediction. To address these limitations, we propose MTUKG, a multi-scale temporal urban knowledge graph dataset. MTUKG represents urban events as time-interval facts, enabling knowledge to be aligned with specific prediction windows. It further introduces a spatial unit-driven method that integrates spatial morphology, semantic distributions, spatial decay effects, and dynamic events to construct temporal functional zones and establish multi-scale temporal associations across micro, meso, and macro urban scales. Experiments on USTP and knowledge reasoning tasks demonstrate that MTUKG significantly enhances predictive performance and provides a structured knowledge resource for modeling the temporal dynamics, functional evolution, and multi-scale dependencies in urban systems. The dataset and source code are available at https://github.com/Jiahaohuhu/MTUKG.

Fuels and petroleum derivatives (e.g., gasoline, diesel, and kerosene), remain essential energy sources for transportation systems, industrial operations, and residential use. Historical data on fuel sales, production, and pricing, typically represented as time series, enable tasks such as demand forecasting, pattern discovery, and anomaly detection, supporting data-driven decision-making in transportation logistics, production planning, and price regulation. However, these efforts are often hindered by the lack of accessible, well-structured datasets. To address this gap, we present StreamFuels, an open-source Python library that serves as a dynamic benchmarking resource for time series mining in the fuel sector. StreamFuels programmatically retrieves the most recent data from official government sources at access time, delivering five curated datasets with over 40,000 time series. The data are cleaned, standardized, and provided in the TSF format, ensuring compatibility with popular time series toolkits. We validate the resource by benchmarking state-of-the-art algorithms across three tasks: forecasting, classification, and pattern analysis (i.e., motif and discord discovery). Our results establish reproducible performance baselines and demonstrate the utility of StreamFuels for advancing research on time series mining and data-driven applications in the fuel sector.

Previous gas-sensing datasets were designed and collected in sealed environments primarily for laboratory research or domain-specific applications. The lack of air-exposed settings limits the study and modeling of the gas properties under natural volatilization. This work bridges this gap by building an unsealed gas-sensing dataset for real-world gas classification, dubbed Daily52, collected from 52 everyday items measured using a self-developed electronic nose (e-nose). We thoroughly benchmark various odor classification models as baselines, among which InceptionTime is found to achieve an impressive 88.24% accuracy in recognizing items. Comprehensive analysis further demonstrates that our dataset is high-resolution and effective for studying the olfactory modality. Potential olfactory research is also discussed beyond our dataset. To our knowledge, this is the first dataset that enables benchmarking methods for classifying real-world gas-sensing data of natural volatiles.

Root cause analysis (RCA) is crucial for enhancing the reliability and performance of complex systems. However, progress in this field has been hindered by the lack of large-scale, open-source datasets tailored for RCA. To bridge this gap, we introduce LEMMA-RCA, a large dataset designed for diverse RCA tasks across multiple domains and modalities. LEMMA-RCA features various real-world fault scenarios from Information Technology (IT) and Operational Technology (OT) systems, including microservices, water distribution, and water treatment systems, with hundreds of system entities involved. We evaluate the performance of six baseline methods on LEMMA-RCA across various settings, including offline and online models, as well as single and multi-modal settings. Our study demonstrates the utility of LEMMA-RCA in facilitating fair evaluation and promoting the development of more robust RCA techniques. The dataset is publicly available at https://www.lemmarca.info.

The meme coin ecosystem has grown into one of the most active yet least observable segments of the cryptocurrency market, characterized by extreme churn and widespread fraudulent behavior. While countless meme coins are deployed across multiple blockchains, they rely heavily on off-chain web and social infrastructure to signal legitimacy. These very signals are largely absent from existing datasets, which are often limited to single-chain data or lack the multimodal artifacts required for comprehensive risk modeling. To address this gap, we introduce MemeChain, a large-scale, open-source, cross-chain dataset comprising 34,988 meme coins across Ethereum, BNB Smart Chain, Solana, and Base. MemeChain integrates on-chain data with off-chain artifacts, including website HTML source code, token logos, and linked social media accounts, enabling multimodal and forensic study of meme coin projects. By providing cross-chain coverage and rich off-chain context, MemeChain serves as a foundational resource for research in financial forensics, multimodal anomaly detection, and automated scam prevention in the meme coin ecosystem.

\begin{abstract} Collegiate running in the United States generates thousands of race results annually in cross country and track and field, yet no large-scale dataset has been publicly available for research. Existing websites such as Athletic.net, MileSplit, and TFRRS host results but do not support bulk download, restricting prior analyses to \textasciitilde{}500 performances, often skewing studies toward male athletes. We introduce the \textbf{National Running Club Database (NRCD)}, the first openly available collegiate running dataset at scale: 143,868 approved performances from 31,351 athletes across 1,423 meets in four sports (cross country (XC), indoor and outdoor track, and road races), 36.2\% women, spanning 2003-2026. Meets from August 2023 onward carry comprehensive course distance, elevation gain and loss, weather at race time, and track venue metadata (99.9\% of XC rows with weather fields). NRCD is \textbf{community-governed} through open submission and expert approval and is maintained as a live database whose meet volume has grown yearly. We release a unified performance standardization framework that operationalizes established distance, elevation, and heat adjustments in one pipeline; \emph{XC-only validation}; heat is a Hadley-band heuristic. \textbf{We recommend gender-stratified modeling}. On XC, full standardization lowers median within-athlete cross-meet variability by 51.1\% (women) and 35.4\% (men) versus raw times. We release the dataset and pipeline with a Python package `nrcd' under FAIR principles, supporting longitudinal athlete modeling, environmental-confounder studies, and gender-equity research in collegiate sport. \end{abstract}

Medical advances are rooted in insights gathered from clinical trials. Technology plays a key role in clinical trial recruitment: patients use specialised platforms to search for clinical trials, and clinical trial recruiters search through patient records to find eligible patients. Existing research shows that personal characteristics of patients influence the outputs of models tasked with medical question answering. While similar classes of neural models are increasingly present in medical information-seeking pipelines, the scientific community lacks resources to investigate how well these models cope with bias in medical retrieval tasks. Here, we present CT-bias, a novel resource for auditing of bias for specific patients and groups of patients where private or sensitive information was recoded as a part of a medical note. To create the dataset, we perturb patient notes from an established patient-trial matching benchmark by inserting specific patient information (sexual orientation, occupation, education level) irrelevant to their medical condition. In an example experiment, we study the impact of this information on the retrieval results of 12 leading open source zero-shot dense retrieval models. We also propose a methodology to inspect models for biases, and observe that, while most models are generally robust to medically irrelevant information, some patient characteristics impact the retrieval results more than others. CT-bias will a useful resource to evaluate future ranking models and CT matching pipelines while the dataset creation protocol itself can be extended to several retrieval and prediction biomedical AI tasks.

Post-Editing (PE) of Machine Translation (MT) output often involves repeating the same lexical and terminological corrections across many segments, especially in specialised and highly repetitive documents. Despite substantial work on Automatic Post-Editing (APE), most available corpora operate at the sentence level, others are synthetic, and overall not designed to study how corrections propagate in realistic Computer-Assisted Translation (CAT) workflows. This paper presents the APEX-VW (Automatic Post-Editing eXperiments on Virtual Wards) Corpus, a new open English-Spanish (EN-ES) dataset built from recent NHS virtual-ward documents and professional PE in Trados Studio, with controlled MT, terminology, and quality assurance settings. The corpus contains seven document-coherent source texts totalling 42k words, translated with four MT systems representing different paradigms and then post-edited by professional translators. Unlike prior resources such as WMT APE corpora, eSCAPE, MLQE-PE, or LangMark, the dataset preserves document order and CAT-tool context, making it suitable for research on terminology normalisation, correction propagation, and human-in-the-loop translation support. The paper describes the corpus design, data preparation, PE setup, and initial corpus statistics, and positions the resource as a benchmark for document-level APE and propagation-aware assistive tools.

Knowledge Distillation (KD) improves the efficiency of recommender systems by transferring knowledge from high-capacity teacher models to lightweight student architectures. Despite its growing popularity, KD research for recommendation remains fragmented, making reproducibility and fair comparison across distillation strategies difficult. We present RecDistillery, the first open-source, model-agnostic framework for standardizing teacher–student KD workflows in recommender systems. The framework provides a modular architecture unifying data preparation, teacher integration, student training, distillation, and evaluation. It supports both native PyTorch-based recommenders and externally trained models, enabling heterogeneous teacher–student distillation within a unified pipeline. RecDistillery facilitates both the reproduction of existing KD methods and the rapid prototyping of new approaches through reusable losses, data loaders, metrics, and configuration modules. We demonstrate its capabilities through representative case studies involving multiple recommendation frameworks and externally imported teachers. The framework code, documentation, and demo examples are publicly available at https://github.com/sisinflab/RecDistillery.

Knowledge Graphs (KGs) enable the integration of data across multiple domains and data types, supporting human interpretability, machine processing, and distributed data access. KGs have become especially valuable for enhancing recommender systems (RSs), which provide personalized user experiences and are now ubiquitous across platforms such as social media, streaming services, e-commerce, news, and search engines. However, reproducibility remains a persistent challenge in developing and evaluating machine learning methods for KG-based RSs, due to heterogeneous data sources and model complexity. Therefore, we introduce FranKGraphBench, a modular benchmarking framework that consolidates baseline models, datasets, and evaluation metrics into a standardized structure. Users can define experiments via configuration files, enabling reproducible sharing of model specifications, datasets, and evaluation settings. FranKGraphBench contains 6 KG-form RS datasets: 3 DBpedia-enriched, 1 Wikidata-enriched, and 2 unenriched. Finally, we validate its effectiveness through a small benchmark and an ablation study to showcase its capabilities and the importance of open settings for reproducible experimentation.

Recommender systems are increasingly expected to combine interaction histories with language. In many cases, a user's next choice may depend on a temporary situation---such as their current mood, plans, or constraints---that is not visible in past behavior. Existing benchmarks, however, mostly evaluate average next-item accuracy, broad instruction following, or content understanding; they do not isolate recommendation episodes in which external situation context is plausibly needed. We introduce \bench{}, a benchmark built on MovieLens for situation-aware recommendation beyond interaction histories. It contains two complementary families of episodes: observed off-trail choices that are real user transitions but hard to infer from prior behavior, and neighbor-supported alternative continuations that represent plausible futures for similar recent histories. Each episode packages a user history, target item, item metadata, and generated life-situation context. The contexts are treated as plausible external circumstances rather than ground-truth user intentions. \bench{} supports controlled evaluation across sequential models, LLM-based recommenders, and hybrid systems by keeping the recommendation episode fixed while exposing or withholding the situation context. By releasing the artifacts, construction pipeline, and evaluation protocol, \bench{} provides a reproducible testbed for studying whether natural-language context improves ranking when interaction history is weak or ambiguous. Benchmark: https://doi.org/10.5281/zenodo.20560655.

Large Language Model personas, explicit profiles specifying a user's attributes, preferences, and behavioural tendencies, are increasingly used to simulate human conversations for user modelling, social reasoning, and behavioural analysis. Evaluating whether such simulations faithfully reflect human conversational behaviour is critical, yet current practice often relies on LLM-as-a-judge approaches that provide limited grounding in observable behaviour and produce opaque scalar scores. We present Eval4Sim, an evaluation framework that measures alignment between simulated and human conversations across three dimensions: adherence, whether persona traits are recoverable from dialogue via dense retrieval; consistency, whether a persona maintains a distinguishable stylistic identity via authorship verification; and naturalness, whether conversations exhibit human-like turn-to-turn flow via dialogue NLI. Unlike optimization-oriented metrics, each dimension takes a human corpus as a reference baseline and penalizes deviations in both directions, distinguishing insufficient persona encoding from over-optimized, unnatural behaviour. The framework is corpus-agnostic: any persona-annotated conversational dataset can serve as the reference. Evaluated over ten simulation corpora, Eval4Sim surfaces systematic trade-offs invisible to single-score methods.

Building Information Modeling (BIM) projects increasingly use Information Delivery Specification (IDS) to formalize information requirements in a machine-checkable XML format. Because IDS conditions are grounded in the Industry Foundation Classes (IFC) vocabulary, authoring them requires expertise in IFC concepts, validation tools, and property set conventions. Existing benchmarks for structured generation do not adequately capture the additional burden of vocabulary conformance and external-validator agreement that IDS imposes. We present Ishigaki-IDS-Bench, the first publicly released benchmark for IDS generation from BIM information requirements. The benchmark contains 166 examples spanning 83 practical scenarios authored in Japanese and English by six BIM/IDS experts, each paired with a gold IDS file and metadata covering input format, turn setting, target IFC versions, and construction domain. Evaluation proceeds in two stages: (i) formal validity scored by the buildingSMART IDSAuditTool along Processability, Structure, and Content, and (ii) content fidelity scored by facet-level macro-F1 against the gold IDS. Across 10 LLMs in zero-shot, the highest Facet F1 is 65.6%, achieved by GPT-5.5, while the highest Content pass rate is only 33.1%, achieved by Claude Opus 4.5. Ishigaki-IDS-Bench is released on Hugging Face (DOI 10.57967/hf/8873) under CC BY 4.0, and the evaluation code is released on Zenodo (DOI 10.5281/zenodo.20550510) under Apache-2.0.

We introduce \dataset{}, a benchmark of 82 textbook-style operations research (OR) problems with bilingual Chinese--English statements, verified numeric answers, and execution-based evaluation. It tests whether a large language model (LLM) can turn natural-language OR descriptions into executable solver code and correct solutions. Baselines distinguish non-reasoning, reasoning, and decomposed-agent systems. The release includes data, evaluation scripts, prompts, and baseline summaries for reproducible comparison.

We present an index lake for Anserini: a remote index representation of all inverted indexes from Anserini's reproducibility suite. Each index is represented as a set of Parquet files on remote object storage, and retrieval is performed through SQL with DuckDB. Through an extensive benchmark of 239 experiments across 50 collections, we empirically confirm our results are consistent with Anserini's. We analyze query latencies over the index lake and conclude that (1) most experiments can be run relatively efficiently, requiring only a couple of seconds per query; and (2) for large collections and learned sparse indexes, queries suffer from increased latency due to high overhead. We are also able to reduce total network transfer, since (1) index lakes are smaller than the corresponding prebuilt Lucene indexes; and (2) up to 90% of the index can be skipped by pushing filters on the postings table to the data lake. The index lake increases reproducibility and makes it easier to run new experiments on standard test collections and prebuilt indexes.

Inverted indexes are a crucial data structure for efficient information retrieval in large text corpora. They enable fast full-text search by mapping each term to the documents in which it appears, on top of which efficient algorithms quickly retrieve the documents relevant to a user query. We present RISE, a novel inverted index library implemented in Rust, designed to deliver high performance and efficiency for information retrieval tasks. RISE leverages Rust's safety and performance to provide a robust solution for building and querying inverted indexes, while offering accessible extensibility through its expressive trait system. While developing RISE, we revisited the inverted-index literature, thereby reproducing numerous prior works using this new test bench. We evaluated RISE against existing libraries, demonstrating competitive query performance across various datasets and workloads, with some algorithms achieving speedups of up to $2\times$ over the current state of the art. Our results indicate that RISE is a promising tool for researchers and practitioners in the field of information retrieval.

Position bias—the systematic effect of where key information is located on retrieval effectiveness—is a critical concern in information retrieval. Although recent studies have identified such bias in dense retrievers, existing analyses focus predominantly on monolingual English, fail to disentangle document length from information position, and lack a standardized framework for controlled evaluation. To address these limitations, we introduce PosIR, a benchmark to systematically diagnose position bias across diverse retrieval scenarios, comprising 310 language–domain dataset combinations spanning 10 languages and 31 domains. Each query is paired with a positive document, a localized reference span, and positional metadata, enabling controlled analysis of retrieval quality across evidence locations and length bins. PosIR provides a complete data generation pipeline with quality-control filtering, along with baseline evaluation scripts. Experiments with 10 widely-used dense retrievers reveal that position bias is pervasive in both multilingual and cross-lingual retrieval, and that it intensifies as documents grow longer. PosIR serves as a foundational resource to advance the development of position-robust retrieval systems.

Node importance estimation is an essential task in network science, with broad applications in recommendation systems and network dismantling. Existing GNN-based methods implicitly operate in a flattened relation space, where diverse interactions are either ignored or collapsed into a single propagation graph. However, real-world networks are far more complex. They comprise multiple types of nodes and edges, and are often multi-layered. As a result, single-layer methods fail to capture the intrinsic and extrinsic interactions among diverse relational categories. This study presents the first dedicated study on node importance estimation in multiplex networks, where a common set of nodes is connected through multiple relational layers. Our objective is to provide a comprehensive analysis by modeling distinct relationships as independent network layers while capturing both intra-layer connections and cross-layer dependencies. The proposed framework, namely HiGH, uses hypergraphs to learn intra-layer structures and capture higher-order interactions and grouping effects by modeling node influences. Specifically, an adaptive importance selector first identifies important nodes at each layer by integrating their local and global information. By constructing hyperedges from selected important nodes rather than from all nodes, the process becomes more targeted for importance-related tasks and reduces the complexity of hyperedge construction. Furthermore, a nested contrastive learning module with inter-layer bridge nodes facilitates cross-layer information transfer, allowing the model to discern node-level universality and specificity across layers. Extensive experiments on three real-world datasets show that HiGH outperforms twelve competing baselines, demonstrating its effectiveness.

State-of-the-art temporal-link-prediction (TLP) models are, in essence, multi-channel information aggregators: they combine an interaction-history channel, a time-encoding channel, and a structure channel. The first two have been refined relentlessly; the structure channel remains a crude afterthought — DyGFormer encodes it as a 1–2-bit neighbour-cooccurrence count. We begin with a measurement: on sparse temporal graphs the classical 1-hop common-neighbour signal is near-random (discriminative AUC ~ 0.50), because two nodes almost never share a direct neighbour; the genuinely discriminative signal lies one hop deeper — the 2-hop cohesive bridge, whose discAUC reaches 0.73–0.98, on both bipartite and non-bipartite graphs. Motivated by this, we propose HAWKEYE, a cohesion-aware structural channel that incrementally maintains the classical k-family of cohesiveness indicators (degree -> k-core -> k-truss) and forms 2-hop cohesive-bridge features. HAWKEYE is a drop-in replacement for a temporal-graph model's native structure channel, with no change to the backbone. Swapping HAWKEYE into DyGFormer improves test AP/MRR over the cooccurrence channel by +0.6 to +10.8 points across six multi-seed-validated datasets (uci, enron, USLegis, CanParl, reddit, mooc; wiki legacy single-seed). On the bipartite recommendation benchmark tgbl-subreddit, a 3-seed single-pass struct-only ablation shows HAWKEYE nearly doubling the baseline test MRR (0.103±0.003 -> 0.204±0.005, +10.1 points across all three seeds), demonstrating large-graph effectiveness; the streaming pipeline scales to the 67M-edge tgbl-flight in five minutes per pass. We further characterise when it helps: the gain tracks a graph's training-free 2-hop discAUC and vanishes on degenerate or saturated graphs — a predictable boundary. All code, data, drawio-editable figures, and figure-generation scripts are released.

Temporal link prediction aims to forecast future connections from historical interactions. As a fundamental problem in temporal graph learning, existing methods exhibit three major limitations. (1) They usually focus on node-level interaction histories and information aggregated from temporal neighbors, while overlooking cohesive information from node-associated subgraphs. (2) They do not sufficiently integrate high-order topological information with temporal information. (3) They are susceptible to confounding factors, which may lead to spurious correlations. To address these limitations, we propose \textbf{CCEC-TGN}, a temporal graph neural network designed for continuous-time dynamic graphs (CTDGs). (1) We leverage continuous coreness to capture time-evolving cohesive information. (2) We use structural encoding, conditional structural interpolation, and gated residual fusion to integrate macro-level structural information with micro-level temporal information. (3) We apply counterfactual consistency regularization to mitigate confounding factors and improve robustness and generalization. Experiments on several real-world temporal graph datasets demonstrate the effectiveness and efficiency of CCEC-TGN, especially in inductive prediction settings. The source code is available at \url{https://github.com/BJHanjz/CCEC-TGN}.

Temporal graph learning has become essential for analyzing real-world systems whose interactions continuously evolve over time, including financial transaction networks, communication systems, and online social platforms. However, learning from large-scale temporal graphs remains computationally challenging when networks are dense and rapidly changing. To address this limitation, we propose a network-curvature-inspired edge sparsification framework for dynamic graph learning. Our proposed method, TRicci, extends classical Forman-Ricci curvature to directed weighted temporal graphs by capturing structural support, temporal recency, and local interaction competition. Experiments on 9 transaction networks and 3 temporal graph benchmark datasets demonstrate that the proposed framework preserves predictive performance across multiple graph-level prediction tasks. The results show that TRicci sparsifies temporal graphs by approximately 80% while reducing end-to-end downstream training and inference time by an average of 55.94%, without substantial degradation in predictive performance. Our findings suggest that temporal curvature can serve as a principled basis for scalable temporal graph learning by preserving predictive temporal-structural information under substantial sparsification.

Short video platforms have revolutionized information consumption but also accelerated the spread of fake news detection. Most existing detection methods use a single-sample paradigm, treating videos as independent instances and ignoring the macro-level event context. To address this issue, this paper proposes an \textbf{E}vent-\textbf{A}ware \textbf{L}abel \textbf{P}ropagation framework over \textbf{H}eterogeneous \textbf{G}raphs for fake news video detection(\textbf{EALP-HG}). The framework constructs an event-centric heterogeneous graph that explicitly models hierarchical relationships among events, videos, and multimodal content, thereby enabling structured reasoning from event context to video evidence. A dual-level attention–based heterogeneous graph attention network is further employed to jointly aggregate node-and type-level information, enabling comprehensive representations for both event and videos. More importantly, EALP-HG introduces an explicit event-aware label propagation strategy, which uses predicted event labels as prior guidance to regularize video predictions. Extensive experiments on two real-world datasets demonstrate the superior performance of EALP-HG. Under 5-fold cross-validation, it outperforms state-of-the-art baselines by 4.73\% and 3.29\% in accuracy, respectively. These results verify the effectiveness of incorporating event-aware supervision into fake news video detection. The code of our work is available at \url{https://github.com/cgao-comp/EALP-HG}.

Heterogeneous graphs (HGs) are widely used to model complex relational data in domains such as traffic networks, biology, and recommendation systems. While heterogeneous graph neural networks (HGNNs) demonstrate strong performance on homophilous benchmarks, they struggle to generalize to heterophilous HGs where connected nodes often hold different labels. Recent spectral HGNNs attempt to address this by modeling complex shift operations by positive non-commutative polynomials, but they suffer from severe over-parameterization and overfitting, especially when data are limited. Furthermore, existing methods often neglect the internal conflict between meta-path subgraphs with diverging homophily ratios. In this paper, we propose \textbf{CLS-HGNN}, a Curriculum Learning enhanced Spectral Heterogeneous Graph Neural Network. To mitigate over-parameterization, we replace dense learnable weight matrices with Jacobi polynomial bases, which decouple the spectral filtering process and significantly reduce complexity while preserving expressiveness. To reduce the impact of conflicts arising from inconsistent homophily, we introduce a curriculum learning strategy that quantifies node difficulty via training loss dynamics and embedding deviations across meta-paths. This allows the model to progressively learn from consistent (easy) patterns to conflicting (hard) ones. Extensive experiments demonstrate that \textbf{CLS-HGNN} achieves state-of-the-art performance on heterophilous HGs and significantly improves training efficiency compared to existing spectral methods.

Risk monitoring tasks like financial fraud detection, money laundering detection, and cybersecurity often involve graph-native datasets that are characterized by extreme class imbalance and evolving structure, as well as deployment constraints that require incremental model updates under bounded temporal memory. Despite recent efforts in graph representation learning, the state-of-the-art Dynamic Graph Neural Networks (DGNNs) have several limitations: They are either designed for class-balanced settings and (some) offer a bounded temporal memory footprint, or are designed for class-imbalanced settings but do not offer a bounded temporal memory footprint. To bridge this gap, we propose HERDGNN: the first DGNN tailored for class-imbalanced settings while keeping a bounded temporal memory footprint. To address class imbalance, HERDGNN integrates deviation-aware regularization pushing abnormal entities away from the distribution of normal ones. To maintain bounded temporal memory, HERDGNN captures temporal information through a compact hierarchical memory whose size is fixed and independent of graph size, enabling live-update training without storing node-level temporal states. We evaluate the effectiveness of HERDGNN on $7$ large-scale real and synthetic financial networks, under a live-update protocol, against $3$ GNN baselines and $5$ state-of-the-art DGNNs. The results show that HERDGNN achieves competitive or superior performance compared to the competing methods, and overall strikes the best balance between classification performance and memory consumption.

Fraudulent user detection in advance is crucial for risk control of real-world applications. One of the primary challenges lies in classifying new users as either benign or fraudulent when only limited user profile information is available. This scenario is often referred to as the Strict Cold Start (SCS) problem. Existing methods generally address this challenge by enhancing the topology or features of SCS nodes. However, these approaches can introduce task-irrelevant noise or require sharing raw cross-scenario features, which is operationally undesirable when scenarios are owned by different teams. To overcome these limitations, we propose a novel graph learning framework called CoMGL. This framework employs a multi-view learning strategy to effectively tackle SCS tasks on graphs. Specifically, the CoMGL regards the interaction bipartite graphs from different scenarios as the multi-views for separately learning user representation, yielding more comprehensive representation for SCS users. Each view encoder consumes only its own scenario-local features, and only the resulting user embeddings are fused, which limits cross-scenario raw feature transfer. Experimental results on real-world datasets demonstrate the effectiveness of the proposed framework compared with the state-of-the-art approaches.

Density-based clustering has been widely used in various applications such as hot spot detection, image segmentation, and geo-social network analysis. DBSCAN, as a representative density-based clustering approach, is known for expensive computation overhead, especially when the data volume is large. Specifically, performing DBSCAN on high-speed continuous incoming data with stringent realtime requirement poses challenges for modern applications, such as traffic monitoring and industry sensor monitoring. This paper proposes RTDBStream, a high-performance DBSCAN approach that utilizes ray tracing cores (RT cores) in modern GPUs for acceleration. RTDBStream transforms DBSCAN in data streams into an efficient ray tracing job in a three-dimensional space. To efficiently utilize the RT cores, two optimization techniques are designed in RTDBStream, namely Identifying Cores by Gridding and Clustering by Creating Mixed-radius Spheres, to significantly accelerate the clustering process. Experimental results show that RTDBStream outperforms existing implementations on both CPUs and GPU CUDA cores, and it achieves up to 16.4$\times$ over existing RT-based implementations.

DBSCAN, one of the most widely used density-based clustering algorithms in both research and industry, has proven to be a powerful tool for discovering complex patterns. Its robustness against noise and outliers has made it a cornerstone of density-based clustering. On large-scale datasets, DBSCAN is costly to compute and, hence, many distributed solutions have been proposed. However, many of them use naive partitioning strategies leading to suboptimal runtimes. To address this limitation, we propose VoronoiSCAN, a novel parallel and exact clustering algorithm that combines the DBSCAN algorithm with Voronoi tessellation for Euclidean distance and low dimensional datasets. By partitioning the input space into cells using Voronoi tessellations, we divide the problem into balanced, local clusterings. The subsequent necessary merging process relies on a novel density connectivity graph construction, which represents the neighborhood of the Voronoi cells. We demonstrate the effectiveness of the proposed algorithm through comprehensive experiments on various datasets, showcasing its potential for large-scale clustering applications and its superior performance compared to other distributed DBSCAN algorithms.

Incomplete multi-view clustering aims to partition unlabeled samples when some views are missing. Recent graph-guided contrastive methods improve clustering by exploiting neighborhood structures from observed views. However, existing methods usually describe local structures with discrete pairwise affinities estimated from incomplete observations. Such affinities are sensitive to view-specific noise, sparse local density, and missing views, where accidental neighbor relations may be transferred across views and further affect contrastive supervision. To address these limitations, we propose Manifold Diffusion Stability-guided Cross-view Graph Supplementation (MDSCGS), a confidence-guided graph supplementation framework for incomplete multi-view clustering. Specifically, MDSCGS estimates manifold diffusion stability from observed view graphs to characterize stable local connectivity beyond direct pairwise proximity, and calibrates this stability through cross-view diffusion-neighbor consistency. Guided by the calibrated stability, high-confidence auxiliary structures are selected to supplement missing target-view relations, while unreliable cross-view relations are suppressed. The augmented view graphs are further integrated into graph-guided contrastive representation learning, enabling the model to exploit complementary cross-view information while preserving stable local manifold geometry. Extensive experiments on four benchmark datasets under different missing rates demonstrate that MDSCGS consistently outperforms representative incomplete multi-view clustering methods.

\begin{abstract} Unsupervised DNA sequence clustering aims to group DNA sequences into biologically meaningful categories based on intrinsic similarities without relying on external annotations. While both global structural patterns and local sequential contextual information is essential for this task, existing methods predominantly rely on structural features, largely overlooking the rich biologically information embedded in the sequential context. To address this limitation, we propose M2DNA, a novel Asymmetric Dual-Stream framework designed for unsupervised clustering. Specifically, M2DNA converts DNA sequences into Frequency Chaos Game Representation (FCGR) images to extract structural priors via a Convolutional Neural Network (CNN), while treating sequences as text to capture local biologically information using a genomic foundation model. These multi-modal features are aggregated via a fusion layer to drive clustering. To mitigate modality laziness in this architecture, we further introduce a Probabilistic Module Dropout (PMD) strategy that compels the learnable branch to extract discriminative semantic features. Extensive experiments on 11 benchmark datasets demonstrate that structural and sequential information are highly complementary, enabling M2DNA to significantly outperform state-of-the-art baselines. To foster reproducibility, our code is publicly released\footnote{https://github.com/Applied-Machine-Learning-Lab/CIKM2026\_M2DNA}. \end{abstract}

Clustering multivariate time series by operational conditions or fault types is important across industrial domains, but existing methods struggle with long, variable-length recordings, localized signatures, and interpretability. We present VISTA (Variable-length Interpretable Scalable Temporal Analysis), an unsupervised time series clustering model for industrial events that jointly captures local and global structure. First, a convolutional encoder embeds each sliding window. Then an eviction-attention mechanism, trained end-to-end via a straight-through estimator, maintains a bounded pool of the top-$P$ most informative embeddings per series, and truncated SVD on this pool produces a fixed-dimensional representation that captures the principal directions of variation across the pooled embeddings of each time series, naturally accommodating variable length. Finally, K-means is applied to all the representations to assign a label to each event time series. Window- and series-level contrastive losses train the pipeline end-to-end. The bounded pool keeps memory usage constant per series and computation linear with sequence length. Intra-pool self-attention scores yield faithful temporal localization (``when''), and a per-cluster channel-attribution score identifies the sensors most relevant to each cluster (``what''). On five industrial datasets, VISTA outperforms all twelve baselines in time series event clustering. Faithfulness tests confirm that these interpretable scores identify windows and channels relevant to VISTA's clustering output and align with domain-expert physical interpretations.

Significant progress has been made in multivariate time series anomaly detection, enabling the effective identification of various anomalies. However, in time sensitive scenarios such as industrial control and microservice monitoring, a critical challenge remains: how to rapidly generate remediation measures that are both interpretable and incur low intervention cost after an anomaly is detected. To close this loop, a system must not only accurately identify anomalies but also trace their root causes. To this end, this paper focuses on algorithmic recourse in noisy multivariate time series and proposes a root cause localization mechanism that integrates causal inference, aiming to precisely identify the core variables that trigger anomalies. Based on the identified root cause variables, counterfactual windows are constructed, from which a downstream recourse generator produces repair suggestions that balance effectiveness with low intervention cost. We introduce a counterfactual based repair framework—CARE (Causal Algorithmic Recourse)—which not only reverses the anomalous time window but also provides causally interpretable remediation strategies. Experimental results on both synthetic and real world datasets demonstrate that CARE achieves precise root cause localization, maintains a high repair success rate, and significantly reduces the intervention cost. This work offers a feasible pathway toward shifting from passive anomaly monitoring to active, interpretable, and autonomous system repair.

The central challenge of one-class tabular anomaly detection lies not only in the absence of anomaly labels, but also in that abnormality can only be defined relative to normality. Such deviations do not have a fixed form in heterogeneous tabular spaces, and thus detection methods have to rely on implicit assumptions about anomaly patterns. For example, reconstruction-based paradigms expect anomalies hard to reconstruct through normal mappings, while distance-based paradigms expect anomalies to exhibit irregularity in local neighborhoods. However, real anomalies do not always follow these assumptions. They may be well reconstructed due to the generalization ability of deep models, or appear locally plausible. Therefore, this paper no longer seeks new reconstruction tricks or local constraints, but instead investigates whether discriminative cues can be discovered from the overall organization of normal patterns themselves. To this end, we analyze the inter-sample connection structure (T-Graph), and find that, under this data-dependent organization, anomalies induce stronger high-frequency responses. This suggests that spectral information captures structural changes induced by anomalous perturbations to normal connection patterns, thereby providing a new structural cue for distinguishing normal and anomalous data. Motivated by this, we propose FASEN (\underline{\textbf{F}}requency-\underline{\textbf{A}}ware \underline{\textbf{S}}pectral \underline{\textbf{E}}ncoding \underline{\textbf{N}}etwork). It first employs wavelet filters to decouple graph signals into multi-scale spectral representations. Subsequently, we design a Frequency Selection module to adaptively learn the frequency preference of normal data, thereby guiding the model to reconstruct utilizing the topological features dominated by normal patterns. Extensive experiments on multiple benchmark datasets demonstrate that FASEN outperforms SOTA methods. Further analysis indicates that this mechanism behaves as a physically interpretable matched filter. The complete code is available at \url{https://github.com/criticcc/FASEN}.

A family of tabular anomaly detectors trains a neural map toward a fixed target under squared-error loss and scores anomalies by the test-time residual; contraction matching, one-step rectified flow, and reconstruction autoencoders all fit this template. We characterize a convergence collapse: better optimization makes the detector worse. At convergence, the learned map tracks the target even off-distribution, so the residual signal vanishes on anomalies as well as on normal data. These detectors therefore rely on implicit non-convergence (early stopping, capacity caps) to retain signal. We argue this is structural: effective anomaly detection requires a locality constraint that blocks unconstrained extrapolation. Classical detectors (kNN, KDE, isolation forests, LOF) enforce locality explicitly; fixed-target neural detectors do not. We formalize the connection by showing that the kernel-regression analog of a fixed-target detector is a finite-bandwidth Nadaraya–Watson smoother, which we call Kernel Contraction Matching (KCM). KCM is closed-form, training-free, and CPU-efficient, yet matches established neural baselines on ADBench. Building on this bridge, we introduce the Kernel-Anchored Regularizer (KAR), which penalizes deviation of the neural prediction from a kernel-weighted average of training targets. Across collapse-prone ADBench datasets and three backbones, KAR mitigates collapse and improves AUROC under prolonged training.

We characterize the attack surface of generative search engines (GSEs) against poisoning attacks in the political domain, from the perspectives of citation selection and personalization. GSEs integrate web search and answer generation with user preferences and backgrounds using large language models (LLMs). They play a crucial role in how users access information on the web. Because anyone can publish content on the web, GSEs are vulnerable to poisoning attacks that manipulate citations to undermine reliable information delivery. Existing studies on citation evaluation focus on how faithfully answers reflect cited content. However, they leave unexamined the two critical aspects to capture the attack surface of GSEs against poisoning attacks: which publishers GSEs prefer to cite, and how personalization affects citation behavior. To fill this gap, we introduce an evaluation framework that characterizes the attack surface of GSEs against poisoning attacks. Our contributions are twofold: (1) we propose a novel metric, \emph{content-injection barrier}, which quantifies the difficulty of injecting arbitrary content onto the web with a given level of publisher authority; and (2) we reveal how personalization affects citation behavior by embedding user profiles into GSEs. We conduct experiments on three major GSEs in the political domain of the United States and Japan. Our results show that (a) the attack surface differs across GSE models; (b) the web search functionality of GSEs shapes the attack surface; (c) ruling parties have a broader attack surface than opposition parties; and (d) user profiles have little influence on the attack surface.

Retrieval-augmented generation (RAG) systems have recently been shown to be vulnerable to fairness and bias manipulation, with prior attacks predominantly targeting the data layer via corpus poisoning. In this work, we expose a fundamentally different and largely unexplored attack surface: the retriever itself. Motivated by the widespread deployment of off-the-shelf neural retrievers as plug-and-play components in modular RAG pipelines, we formalize the Adversarial Fairness Attack against RAG (AFA-RAG), where an adversary manipulates only retriever parameters, leaving both the corpus and generator untouched, to systematically skew evidence exposure toward a target demographic while preserving semantic relevance and stealth. To instantiate this threat, we propose AFA-RL, a reinforcement learning framework that treats the retriever as a strategic ranking agent interacting with a RAG environment. AFA-RL learns adversarial ranking policies that subtly reshape retrieved evidence to bias downstream response generation, guided by a fairness-aware reward with strict relevance and distributional constraints. Experiments across multiple datasets and large language model backbones show that AFA-RL can cause disproportionate fairness degradation with negligible impact on standard retrieval metrics. Our results uncover a previously overlooked supply-chain vulnerability in modular RAG systems and underscore the urgent need for robustness guarantees at the retriever level.

Multimodal Large Language Models (MLLMs) are increasingly deployed in high-stakes domains where fairness is a critical safety requirement. In practice, these models are continually updated through continual learning (CL) to adapt to evolving tasks and data distributions. Prior work has shown that backdoor attacks can manipulate MLLM responses through hidden triggers, but naively implanted backdoors degrade as models undergo subsequent updates of CL. Although fairness has emerged as a central concern for MLLM deployment, whether backdoor-induced fairness violations can survive CL remains unexplored, leaving two critical questions unanswered: (1) whether a backdoor can reliably induce fairness violations in MLLMs, and (2) whether such fairness-targeted backdoors can persist through continual learning. We bridge this gap by proposing \textit{Persistent Fairness Backdoor Attack (PFBA)} to inject persistent and group-specific discrimination into MLLMs. Specifically, PFBA achieves this through two novel mechanisms. The \textit{Latent Space Fairness Reinforcement} reshapes the model's deep feature geometry by anchoring privileged-group representations to preserve utility while repelling and clustering targeted-group representations to sustain discrimination, and the \textit{Continual Learning Simulation} iteratively optimizes the trigger against simulated parameter drift to ensure backdoor persistence across future updates. Extensive experiments demonstrate that PFBA induces severe fairness disparities that persist across continual learning rounds, evading standard backdoor defenses. The data and code are publicly available at \href{https://github.com/lyygua/Persistent-Fairness-Backdoor-Attack-LVLM.git}{https://github.com/lyygua/PFBA}.

Multivariate time series (MTS) prediction is widely used in safety-critical domains such as traffic management, energy systems, weather forecasting, and financial analysis. Recent studies have shown that MTS prediction models are vulnerable to backdoor attacks, where poisoned training samples with hidden triggers can manipulate model predictions at run time. However, existing backdoor attacks on MTS prediction mainly focus on a single target pattern, forcing attacked variables to follow the same predefined trend. This setting overlooks a more realistic and challenging threat: in real-world multivariate systems, different variables are often causally correlated, sometimes positively and sometimes negatively, so an attacker may need to manipulate multiple target variables toward different target patterns simultaneously while keep the predictions normal for unpoisoned data. To address this gap, we propose MTBTime, a backdoor attack framework for multi-objective target patterns in MTS prediction. MTBTime first identifies vulnerable poisoning positions by combining Granger causality, neural causality scores, and prediction-error-based screening. It then constructs an adaptive causal graph convolutional trigger generator to learn inter-variable dependencies and optimize stealthy triggers that can jointly manipulate multiple target variables toward their respective target patterns. Extensive experiments on multiple benchmark datasets and mainstream MTS prediction models show that MTBTime achieves stronger attack effectiveness than existing methods in multi-target scenarios, while maintaining clean prediction performance and attack stealthiness. These results reveal a previously underexplored vulnerability of MTS prediction models under multi-objective backdoor attacks.

Adversarial attacks pose a major threat to link prediction models. Carefully chosen triple-level perturbations can exploit weaknesses in the learned representations and change the model's ranking decisions. Despite their importance, adversarial attacks on knowledge graph embeddings (KGE) models remain less studied than their counterparts in areas such as image classification. One reason is that KGE inputs are discrete relational triples, not continuous pixel values. Moreover, KGE scores are computed within multi-relational prediction settings, where each relation can produce a different score distribution. As a result, score-based attack strategies cannot be directly transferred to KGE models. Studying training-data perturbations therefore both exposes vulnerabilities and identifies informative graph edits that can improve performance. In this work, we use adversarial triple modifications as controlled probes of KGE decision behavior. We introduce a decision-margin perspective that measures how clearly each training triple is ranked above its strongest negative alternatives. Our results reveal systematic vulnerability patterns, as well as improvement opportunities, that are shared across different KGE models. At the same time, these effects are not determined by the embedding model alone. The impact of training data perturbations also depends on the structure of the underlying knowledge graph. Overall, our study reframes adversarial probes as diagnostic instruments for KGE models.

Graph pre-training and prompt learning paradigm have recently achieved remarkable success by improving the adaptability of Graph Neural Networks (GNNs) and alleviating the reliance on large task-specific datasets. However, we find that this paradigm inadvertently increases the vulnerability of GNNs to model extraction, a type of attack in which adversaries replicate a target model by observing its outputs. A particularly challenging and realistic scenario arises when attackers cannot access any target training data, known as the data-free setting. To address this, we propose $\text{GrAP}^3$, a novel framework that enables the reconstruction of high-fidelity surrogate models under this resource-constrained condition. We first conduct a formal risk analysis to examine the impact of pre-training and prompt learning on model extractability. The analysis shows that these paradigms jointly tighten the theoretical upper bound of surrogate–target discrepancy. They are achieved by reducing representation, classification, and distribution gaps, which is the barrier to model extraction. Building on this insight, $\text{GrAP}^3$ generates structurally coherent and label-aware query graphs using a multi-class contextual stochastic block model. It further integrates a learnable prompt subgraph whose embeddings and structural connections are jointly optimized with a pre-trained surrogate through alternating optimization. Comprehensive experiments on six benchmark datasets demonstrate the superior model extraction performance of $\text{GrAP}^3$ without any access to the training data.

Graph Neural Networks (GNNs) have achieved remarkable success across diverse applications, yet they remain highly vulnerable to adversarial attacks that maliciously perturb graph structure. Existing defenses often lack rigorous theoretical grounding, rely on attack-specific heuristics, or require costly retraining procedures such as adversarial training. To address these limitations, we propose Kernel-Complexity Edge Sanitization (KCES), a training-free and model-agnostic framework for defending against structural attacks. KCES is built upon Graph Kernel Complexity (GKC), a principled metric derived from the graph Gram matrix that appears in a generalization upper bound on the GNN test error. From this bound, we define an edge-specific KC score that quantifies each edge’s structural influence via its induced change in GKC. KCES then identifies and prunes high-KC edges, which are empirically enriched with adversarial perturbations under structural attacks, to mitigate their harmful impact. Computationally efficient and scalable, KCES operates as a lightweight preprocessing step without retraining and can be seamlessly integrated with existing defenses. Extensive experiments demonstrate that KCES consistently outperforms representative robust baselines across diverse attack settings and scales effectively to large graphs. Supported by theoretical analysis and extensive empirical validation, KCES provides a principled and efficient framework for securing GNNs. Our code is available at https://github.com/karpning/KCScore.

Model inversion attacks (MIAs) aim to reconstruct high-quality copies of inaccessible private training data from deep neural networks, thereby posing a severe threat to data privacy. Training-based MIAs are only applicable to reconstructing low-resolution grayscale images. Optimization-based MIAs suffer from complex training of attack models, reconstruction quality degradation caused by iterative optimization, and high query and time overhead. Although diffusion-based MIAs achieve significant advantages in reconstruction quality and generation efficiency, this paradigm still faces two major limitations: 1) Unfiltered public datasets, which severely constrains reconstruction quality. 2) Low execution efficiency, particularly the substantial computational and time overhead incurred by training conditional diffusion models from scratch in the image space. Thus, we propose Lightweight Latent Diffusion Fine-tuning for Model Inversion Attack (LLDF-MI), a novel method that fine-tunes the pre-trained latent diffusion models (LDMs) via Low-Rank Adaptation (LoRA) to achieve efficient and high-quality black-box MIA. Specifically, we first constructs a Top-K positive face fine-tuning dataset via head pose detection and target model confidence filtering for precise visual-identity mappings. On this basis, we design a Multi-Identity LoRA module with a "partially identity-specific + partially shared" parameter allocation strategy, combined with conditional encoding and substitute-based identity loss, to precisely guide the generated images toward the target identity feature space. Experiments show that its attack performance significantly outperforms existing black-box and label-only MIAs, and is comparable to advanced white-box MIAs. Furthermore, it substantially reduces the training time overhead of the attack model and effectively improves the generation efficiency of attack images.

Re-ranking plays a crucial role in modern information search systems by refining the ranking of initial search results to better satisfy user information needs. However, existing methods show two notable limitations in improving user search satisfaction: inadequate modeling of multifaceted user intents and neglect of rich side information such as visual perception signals. To address these challenges, we propose the Rich-Media Re-Ranker framework, which aims to enhance user search satisfaction through multi-dimensional and fine-grained modeling. Our approach begins with a Query Planner that analyzes the sequence of query refinements within a session, decomposing the query into clear and complementary subqueries to enable broader coverage of users’ potential intents. Subsequently, moving beyond primary text content, we integrate richer side information of candidate results, including signals modeling visual content generated by the VLM-based evaluator. These comprehensive signals are then processed alongside carefully designed re-ranking principle that considers multiple facets, including content relevance and quality, information gain, information novelty, and the visual presentation of cover images. Then, the LLM-based re-ranker performs the holistic evaluation based on these principles and integrated signals. To enhance the scenario adaptability of the VLM-based evaluator and the LLM-based re-ranker, we further enhance their capabilities through multi-task reinforcement learning.

As large-scale visual-document corpora such as arXiv papers and enterprise PDFs continue to grow, visual-document retrieval has gained attention; yet it still lacks a deployable system that lexically indexes visual documents to serve queries without neural encoding at scale. Existing methods either achieve strong retrieval quality with VLM-based dense or multi-vector models but require neural query encoding at serving time, or avoid query encoding with OCR- or caption-based BM25 at the cost of time-consuming text extraction or generation. To fill this missing serving regime, we present V-SPLADE, an inference-free sparse retriever for visual-document retrieval. However, such inference-free multimodal learned sparse retrieval systems remain underexplored and have not yet shown dense-level effectiveness under high sparsity. We attribute this limitation to a lexical grounding problem: visual sparse representations often fail to capture the lexical content embedded in document images. To address this, we introduce caption-gated token supervision, a training-only signal that uses VLM-generated captions as lexical cues to activate retrieval-relevant vocabulary dimensions. With this supervision, V-SPLADE improves average NDCG@5 across six benchmarks by +13.7pp over the same-scale dense baseline and by up to +6.2pp over OCR- or caption-based BM25 baselines. On an 18.7M-document corpus, it doubles R@5 over the same-scale dense baseline and further improves competing retrievers through score fusion by up to +1.6pp R@5. Code and model weights are available at github.com/naver/v-splade and huggingface.co/naver/v-splade-quality, respectively.

Knowledge-based Visual Question Answering (KB-VQA) aims to answer queries that necessitate reasoning over external knowledge sources beyond the visual content. Typically, current methods fuse multimodal features to retrieve external information, subsequently leveraging Multimodal Large Language Models (MLLMs) to derive answers from the retrieved evidence. However, these methods often struggle to capture structural associations within complex contexts to effectively filter noise. Furthermore, they frequently fail to ensure that the reasoning process remains strictly faithful to the retrieved evidence. To address these challenges, we propose SAFE-G, a Structure-Aware Faithful Evidence-guided Generation framework, which enables precise evidence localization and trustworthy reasoning. Specifically, we first employ a coarse-grained hybrid search fusing visual and textual modalities to recall candidate documents, and subsequently implement a structure-aware fine-grained graph retrieval that captures structural dependencies to filter noise and pinpoint precise evidence. Moreover, we introduce a reinforcement learning (RL) strategy with an evidence-grounded reward that assigns credit to correct answers only when the selected evidence is correct. This strict alignment constraint compels the model to anchor its response in the retrieved context, effectively enhancing its capability to locate evidence via multimodal features and perform faithful reasoning. Extensive experiments on the Encyclopedic-VQA and InfoSeek benchmarks demonstrate that SAFE-G outperforms prior methods by a margin of 8.9% and 3.5%, substantially enhancing the overall reasoning accuracy. Our source code is publicly available at: https://github.com/MINE-USTC/SAFE-G.

Sketch-based 3D shape retrieval (SBSR) aims to retrieve 3D shapes that are consistent with the category of the input hand-drawn sketch. The core challenge of this task lies in two aspects: existing methods typically employ simplified aggregation strategies for independently encoded 3D multi-view features, which ignore the geometric relationships between views and multi-level details, resulting in weak 3D representation. Simultaneously, traditional SBSR methods are constrained by visible category limitations, leading to poor performance in zero-shot scenarios. To address these challenges, we propose \textbf{Multi-View Hierarchical Graph Neural Network} (MV-HGNN), a novel framework for SBSR. Specifically, we construct a view-level graph and capture adjacent geometric dependencies and cross-view message passing via local graph convolution and global attention. A view selector is further introduced to perform hierarchical graph coarsening, enabling a progressively larger receptive field for graph convolution and mitigating the interference of redundant views, which leads to more discriminative hierarchical 3D representation. To enable category-agnostic alignment and mitigate overfitting to seen classes, we leverage CLIP text embeddings as semantic prototypes and project both sketch and 3D features into a shared semantic space. We use a two-stage training strategy for category-level retrieval and a one-stage strategy for zero-shot retrieval under the same model architecture. Under both category-level and zero-shot settings, extensive experiments on three public benchmarks demonstrate that MV-HGNN outperforms state-of-the-art methods.

Universal Cross-Domain Retrieval (UCDR) requires models to generalize to unseen domains and unseen categories at inference time. Recent prompt-learning methods for vision--language models have achieved strong performance on object-centric benchmarks, yet their behavior under cluttered multi-object scenes remains under-explored. In such scenes, co-occurring objects can act as contextual distractors and induce \textbf{prompt contamination}: visually salient distractors may dominate the candidate tokens used for instance-adaptive prompt generation, causing retrieval representations to drift away from the designated primary target. To diagnose this gap, we construct \textbf{DomainCOCO}, a Realistic UCDR benchmark with 180,000 images across six domains, where each sample contains a designated primary category and contextual distractors. DomainCOCO is intended as a controlled diagnostic benchmark for studying target-aware retrieval under multi-object clutter and stylized domain shifts. We further provide \textbf{Text-Guided Rectification Network (TGR-Net)}, a practical rectification baseline that uses text embeddings as semantic guidance to re-weight patch-level visual evidence before prompt generation. Its Candidate Prompt Purification (CPP) module reduces distractor-dominated prompt synthesis, while a teacher--student strategy with a learnable null token enables label-free inference. Multi-scale feature fusion and gallery-side refinement are used as optional inference components for accuracy-oriented deployment. Experiments on DomainNet and DomainCOCO show that TGR-Net provides consistent gains over strong CLIP-based baselines, with larger gains in the more cluttered DomainCOCO setting. Code is available at \url{https://github.com/YUEcjy13/TGR-Net}.

Cross-modal retrieval aims to identify semantically relevant instances from a gallery of another modality given queries of interest. Existing methods typically assume that online queries follow the same distribution as the source domain. However, due to the complexity and diversity of real-world queries, online query streams often substantially deviate from the source distribution; such deviations distort the geometry of the shared representation space and degrade retrieval performance. In this paper, we demonstrate that distribution shifts are inherently anisotropic within the representation space, where different geometric directions exhibit varying levels of semantic sensitivity and shift intensity. Leveraging this insight, we propose \textbf{P}rojected \textbf{R}epresentation via \textbf{I}nduced \textbf{S}pectral \textbf{M}etric (\textbf{PRISM}), which employs spectral decomposition to separate semantic-sensitive components from shift-intensive variations. During test-time adaptation, PRISM selectively confines adaptation updates to the subspace associated with high shift intensity but low semantic sensitivity, thereby suppressing distributional perturbations while preserving semantic consistency. We further show that PRISM achieves a tighter generalization bound under distribution shifts. Extensive experiments on six benchmarks demonstrate the effectiveness of PRISM across diverse distribution-shift scenarios. Code is available at: \textcolor{magenta}{\url{https://github.com/W2Wzj/PRISM}}.

Large language models are increasingly used to search biomedical knowledge bases. In this setting, queries describe the desired biomedical entity in natural language, and the system retrieves matching candidates. Existing methods typically represent candidate entities through text, leaving out available structural information such as molecular structure or protein sequence. Text descriptions can omit fine-grained structural information, making it difficult for text-only representations to distinguish candidates with similar descriptions but different structures. A natural idea is to introduce structural modalities. However, structural and textual representations of the same entity can encode partially overlapping information, creating cross-modal redundancy. Naive fusion can then reinforce shared information and reduce the contribution of additional structural signals. We present OrthoMol, a multi-modal framework for biomedical entity retrieval whose entity-side fusion explicitly accounts for this redundancy. Its core step, Residual Orthogonal Projection (ROP), decomposes the structural embedding into two parts: a text-aligned component and an orthogonal residual. ROP then removes the text-aligned component and uses the orthogonal residual to augment the text embedding as a complementary structural signal. We evaluate OrthoMol on STaRK-Prime and ProtDescribe, two biomedical retrieval benchmarks in which natural-language queries are matched against candidate entities described in text. On both benchmarks, OrthoMol outperforms foundation-model retrieval baselines and improves over the text-only version of the same backbone. Notably, on ProtDescribe, it reduces the mean rank of the target protein by 29\% relative to the text-only backbone.

Retrieval-Augmented In-Context Learning (ICL) faces fundamental Information Retrieval (IR) bottlenecks when applied to continuous graph-structured data (e.g., molecules), as 1D textual representations (e.g., SMILES) fail to capture latent topological causalities. This representational bottleneck becomes catastrophic in \textit{Activity Cliff} scenarios, where minor structural modifications trigger extreme target value divergences. In such environments, standard retrieval methods inject contradictory exemplars, causing LLMs to suffer from the Parroting effect---blindly mimicking the labels of retrieved neighbors rather than discerning true structural causality. Furthermore, existing retrieval strategies apply a uniform criterion regardless of the task nature, neglecting the distinction between local structural motifs and global physicochemical properties, thereby failing to provide optimal inference contexts. To address these challenges, we propose $\boldsymbol{S^3}$\textbf{-Prompt}, a novel framework integrating three key strategies: \textbf{Structural Decomposition} to preserve structural integrity via scaffold and side-chain soft embeddings; \textbf{Structural Contrastive Learning} to sharpen decision boundaries at inference time via an activity-aware contrastive loss; and \textbf{Strategic Retrieval} to employ task-adaptive retrieval modalities (topological vs. semantic) tailored to specific biochemical characteristics. To rigorously evaluate this, we introduce a \textbf{Conflicting Context Subset} to expose the vulnerabilities of the parroting effect. Experimental results demonstrate that $S^3$-Prompt achieves state-of-the-art performance (e.g. up to 2.8\% improvement in ROC-AUC) on MoleculeNet classification tasks while maintaining highly competitive results and consistent architectural synergies across regression tasks. Notably, it exhibits significant improvements in prediction robustness within challenging activity cliff regions.

Dense retrieval systems are core components of modern search engines, recommendation platforms, and retrieval-augmented generation pipelines. They encode documents and queries into dense embeddings, enabling efficient semantic search via vector similarity. However, because document embeddings are computed offline and stored in static indexes, these systems struggle to adapt to user feedback or evolving search intent. To address this limitation, we introduce \emph{embedding surgery}, a lightweight approach for adaptive ranking correction in dense retrieval. The method applies localized, minimal updates to selected document embeddings at query time, guided by editorial feedback, user interactions, or pseudo-labels from large language models. We formulate embedding surgery as a convex optimization problem that enforces ranking constraints while minimizing modifications to the affected document representations. We integrate embedding surgery into standard dense retrieval pipelines and evaluate it on TREC Deep Learning, TREC Robust, TREC CAsT, and MS MARCO benchmarks. Results show consistent improvements (e.g., up to +60.64\% relative improvement in nDCG@10 on DL-Hard under editorial feedback), even under noisy feedback, with low computational cost and without disrupting the global structure of the embedding space. Extensive experiments show that ranking corrections propagate to semantically related queries and that embedding updates can be applied safely and efficiently to scalable Approximate Nearest Neighbor indexes via simple in-place overwriting, without requiring costly index reconstruction. Finally, embedding surgery complements query adaptation methods such as CoRocchio, yielding additional gains while being more robust to noisy feedback.

Cross-encoders fine-tuned from Transformer backbones remain the standard for second-stage re-ranking, and recent knowledge-distillation strategies have closed much of the gap with LLM re-rankers. However, these strategies have not been compared under controlled conditions. In particular, it remains unclear how distillation from LLM rankers compares to distillation from strong cross-encoder teachers, or to purely supervised objectives. It is also unclear how much newer backbones (RoBERTa, ELECTRA, DeBERTaV3, ModernBERT) contribute compared to the original BERT. We run 162 controlled training runs (9 backbones $\times$ 6 objectives $\times$ 3 seeds), spanning pointwise, pairwise, and listwise losses with both human labels and two distillation signals, and evaluate on TREC-DL, MS MARCO dev, BEIR, LoTTE, and Robust04. We find that objectives emphasizing relative comparisons—pairwise MarginMSE and listwise InfoNCE—consistently outperform alternative objectives, including more complex listwise LLM distillation, across all backbones, and switching objective yields gains comparable to moving up one backbone size tier. A controlled disentanglement further shows that, once the negative-sampling pool is matched, even a simple pairwise Hinge loss with ColBERTv2 hard negatives matches---and on out of domain beats---listwise LLM distillation, indicating that the quality of the negatives is at least as important as the choice of loss. We release all 162 trained models on HuggingFace and a unified training codebase.

Large-scale supervision is essential for training modern ranking models, yet obtaining high-quality human relevance annotations remains costly. Click data has long been used as a low-cost alternative, and recent advances in large language models (LLMs) have introduced LLM-based relevance annotation as another promising source of supervision. This paper investigates whether LLM annotations can replace click data for learning to rank (LTR) through a comprehensive comparison across multiple dimensions. Experiments on both a public dataset, TianGong-ST, and an industrial dataset, Baidu-Click, show that click-supervised models perform better on high-frequency queries, whereas LLM-annotation-supervised models are more effective on middle- and low-frequency queries. Further analysis suggests that click-supervised models better capture document-level signals, such as authority and quality, while LLM-annotation-supervised models are more effective at modeling semantic matching between queries and documents and distinguishing relevant from non-relevant documents. Motivated by these findings, we explore two training strategies, annotation-aware data scheduling and frequency-aware multi-objective learning, to integrate both supervision signals. Both strategies improve ranking performance across queries at all frequency levels, with frequency-aware multi-objective learning achieving stronger results. Our code is available at \url{https://github.com/Trustworthy-Information-Access/LLMAnn_Click}.

Whole-page optimization (WPO) decides how search and recommendation results are surfaced to users, and large language models (LLMs) open a new route to it by treating page generation as sequence generation. Adapting LLMs to web-scale WPO, however, remains bottlenecked by the need for costly human annotations and by the mismatched granularity between page-level coherence and item-level placement. In this work we show that these two challenges are coupled: implicit user feedback alone suffices for alignment, provided the reward signal is decoupled into two complementary granularities. We propose PageLLM, a reward-based fine-tuning framework that (i) turns implicit feedback into four contrastive preference-pair families covering relevance, ranking, diversity, and redundancy, (ii) learns a coarse page-level reward and a fine item-level reward that captures engagement-sensitive position swaps, and (iii) combines both rewards in PPO-based RLHF over a pre-trained LLM. Extensive experiments on seven Amazon categories against eleven baselines show that neither reward alone is sufficient---dropping the page-level or item-level signal reduces NDCG@100 by 17.8% and 15.2% respectively, whereas the joint reward improves NDCG@100 by up to 46.8%. Deployed in a 10M-user online A/B test, PageLLM raises GMV by 0.44% and click-through rate by 0.14\%, confirming that multi-grained rewards from implicit feedback scale to production WPO.

Reranking is instrumental in multi-stage recommender systems as it produces the target sequence of items exposed to users. Generative models have shown promise in this field recently, surpassing traditional discriminative paradigms. However, current methods generally follow an item-level generation manner. They either construct a sequence by irreversible autoregression, or iteratively refine a given sequence through item swapping and replacements. Such item-level decision horizons may cause the solution to be trapped in a local optimum, and also limit the ability to capture dependencies within the sequence, thus failing to pursue a global optimal objective. Furthermore, these methods mainly rely on historical contexts for modeling. We argue that in addition to such "past descriptors", the generation of a target sequence also requires directional guidance related to the future. In this paper, we propose \textbf{SGeture} (\textbf{S}elf-\textbf{G}uided \textbf{Ge}nerative Reranking via Pic\textbf{tur}ing the P\textbf{e}rmutation), a sequence generation framework to address the above limitations. First, a self-guidance mechanism is designed to shape a future "guidance": (i) we enable the model to learn the correspondence between an exposed sequence and its \textit{own} resulting future; (ii) at inference, the model is then equipped to self-control the future to be a \textit{desired} one, therefore self-guiding the generation of a target sequence. Second, we present a holistic sequence generator for reranking. Specifically, a diffusion module is employed to progressively "picture" the permutation over candidate items in a sequence-level manner instructed by the above guidance, thereby avoiding local decision traps. Extensive offline experiments on public datasets and online A/B tests across different platforms of a real-world video application justify that our framework consistently outperforms existing generative approaches in reranking.

As a critical stage in the modern multi-stage recommendation pipeline, list-wise reranking determines the final exposed item sequence and directly shapes the user experience by modeling complex dependencies among candidate items within a list. Existing methods typically formulate this task as selecting indices from the local input list. However, this approach suffers from a semantically inconsistent action space: the same output neuron (logit) corresponds to different items across different samples, preventing the model from forming a stable, item-intrinsic representation. To address this, we propose GloRank (Global Action Space Ranker), a generative framework that shifts reranking from selecting local indices to generating global identifiers. Specifically, we represent items as sequences of discrete tokens and reformulate reranking as a token generation task. This design effectively decouples the scoring mechanism from the order of variable inputs, ensuring that items are evaluated against a consistent global standard. To further strengthen the framework, we adopt a two-stage optimization pipeline: supervised pre-training to initialize the model with high-quality behavioral demonstrations, followed by reinforcement learning-based post-training to directly maximize list-wise utility. Experiments on two public benchmarks and a large-scale industrial dataset, complemented by online A/B testing, demonstrate that GloRank consistently outperforms state-of-the-art baselines and also achieves strong robustness in cold-start scenarios. GloRank has been deployed in production on our large-scale short-video recommendation platform, serving over a hundred million users daily.

Modern industrial recommender systems have increasingly adopted the Generator-Evaluator (G-E) framework for the re-ranking stage. Within this paradigm, the generator produces candidate item lists from a pool filtered by upstream retrieval and ranking modules, while the evaluator scores these lists and selects the highest-scoring one for final exposure per request. However, on sequential platforms (e.g., short-video apps), users consume items continuously, ignoring artificial list boundaries. Conventional evaluators score lists by aggregating point-wise values, implicitly assuming exposure independence. This fails to capture critical session-level dynamics, such as contextual dependencies, user continuation, and diminishing marginal utility from repetitive content. To bridge this gap, we propose \textbf{SWIM} (\textbf{S}tep-\textbf{W}ise \textbf{I}ntegrated \textbf{M}easure), a list-level evaluator that models user behaviors as a finite-horizon prefix session-level survival process. SWIM estimates the prefix-conditioned contribution of the current list to the session-level objective by factorizing it into a recursive survival distribution and reached-position conditional rewards. Leveraging a causally-masked Transformer, SWIM efficiently estimates continuation probabilities and utilities in parallel, satisfying strict industrial latency constraints. Extensive experiments demonstrate that SWIM significantly outperforms baselines in listwise reranking tasks, yielding substantial improvements in overall recommendation engagement.

Stock ranking under non-stationary market conditions requires representations of inter-asset dependencies that adapt to changing market states. Prior work includes methods that learn dynamic stock--stock relations from price-based representations, as well as methods that incorporate market context into stock representations. We present HSCA (Heterogeneous Stock Conditioned Attention), which explicitly models the predictive affinity space. HSCA combines state-conditioned query modulation, State Anchors, and a supervised residual topology calibrator with Alpha-Guard stabilization. A State Gate conditions stock queries on the current market state, while State Anchors introduce global state references into the attention bank. During training, a Structural Hinge Loss supervises pre-softmax affinities using pairwise targets derived from cross-sectionally standardized forward returns. The loss penalizes violations of a margin that ranks directionally consistent peers above conflicting peers, whose directed stock pairs are identified as Toxic Edges. Experiments on CSI300 and CSI800 under a common evaluation protocol show consistent ranking and portfolio gains. Component and sensitivity analyses distinguish the roles of the proposed mechanisms. Regime-conditioned evaluation shows that residual topology calibration improves peer ordering and reduces conflicting exposure relative to the base peer topology in every decoded market state. The largest peer-margin and ordering gains and conflict reductions occur in bear regimes. At inference time, removing the learned residual topology correction or shuffling it within each source row produces lower mean IC and RankIC and higher mean Toxic Attention Mass than the learned correction in both universes.

Most recommender systems remain fundamentally dependent on user–item interaction data for training. Recent studies suggest that LLMs can be used either as zero-shot recommenders or as user simulators, but whether LLM-generated signals can substitute for missing behavior data and train a recommender system remains an open question. We study Catalog-to-Supervision Training for recommendation: the model is trained using only the item catalog, with no user–item interaction logs or popularity statistics. At inference, the model receives the current user's observed short history as the query, exactly as in standard sequential recommendation. We propose CASTRec (CAtalog-to-Supervision Training for launch-time Recommendation), a catalog-to-supervision training framework that uses an offline LLM to generate supervised preference signals. Given only the catalog, CASTRec synthesizes catalog-grounded synthetic preference trajectories, grounds them in the actual catalog items, and trains a lightweight encoder for launch-time recommendation. Across benchmarks, CASTRec outperforms baselines, despite using zero interactions during training. It is especially effective in short-history and long-tail regimes, where interaction-trained models are most limited.

Multi-behavior recommendation (MBR) leverages auxiliary behavioral signals, such as clicks and add-to-cart, to enhance target behavior prediction like purchases. While recent graph neural network–based approaches have achieved strong performance by systematically propagating auxiliary behavior signals, they still suffer from two fundamental challenges inherent to auxiliary behaviors: (1) missing auxiliary signals, which hinder generalization to items without auxiliary observations, and (2) unreliable auxiliary signals, which amplify noise misaligned with the target behavior. To address these challenges in a unified manner, we propose \proposed, an environment-conditioned MBR framework that addresses missing and unreliable auxiliary signals through two complementary modules conditioned on auxiliary observability. Extensive experiments demonstrate that \proposed consistently outperforms state-of-the-art baselines, achieving up to 7.82\% gains in HR@10 overall and up to 44.2\% gains for target items without auxiliary observations, highlighting its ability to capture hidden preferences beyond observed auxiliary relations. Our code is available at: \url{https://github.com/LSH0411/BOAR}.

Multi-Domain Recommendation (MDR) aims to model user behaviors across multiple domains with heterogeneous distributions while balancing knowledge sharing and domain-specific specialization. Mixture-of-Experts-style (MoE-style) architectures have become a dominant paradigm for MDR, but they often suffer from expert load collapse, where the routing network overuses a few experts and leaves others under-trained, thereby reducing effective model capacity and scalability. To address this issue, we introduce Dynamic Hyper-Routing (DHR), a hypernetwork-inspired routing mechanism that generates sample-conditioned routing parameters and alleviates expert load collapse without relying on sensitive load-balancing coefficients. More importantly, DHR reveals a mathematical isomorphism between MoE-style routing and Transformer self-attention when experts are viewed as tokens, providing a unified perspective for bridging expert modeling and Transformer-style scalable computation. Building on this insight, we propose MoEFormer, a unified architecture for MDR that reformulates expert-based domain specialization into a Transformer-style computation paradigm. Specifically, the MoEFormer consists of three key designs: (1) A Multi-Level Tokenization that converts heterogeneous multi-domain inputs into dimension-aligned tokens at the field, context, reasoning, and domain levels. (2) A Domain-Disentangled Mask that constrains token interactions to prevent negative transfer and enforce domain-aware information flow. (3) A Domain-Adaptive Transduction Unit that facilitates multi-level token interactions while balancing domain knowledge sharing and specialization. Extensive experiments on three large-scale offline datasets and one real-world online dataset demonstrate the effectiveness and scalability of the MoEFormer. Moreover, the deployment of MoEFormer in a commercial online multi-domain advertising system achieves a maximum conversion rate improvement of 6.37%

\begin{abstract} While Graph Neural Networks (GNNs) have established themselves as the state-of-the-art paradigm in recommender systems, their performance is often hampered by severe data sparsity. Cross-domain recommendation (CDR) offers a promising solution by leveraging multi-domain interactions. However, existing methods typically rely on specialized, end-to-end architectures. This isolation prevents the reuse of mature single-domain GNNs. Moreover, these methods typically employ static or uniform transfer mechanisms, failing to capture the \emph{interest heterogeneity across domains} and the \emph{personalized nature of cross-domain correlations}. To bridge this gap, we propose \textbf{GPCDR}, a general \textbf{G}raph \textbf{P}rompting framework for \textbf{C}ross-\textbf{D}omain \textbf{R}ecommendation. Designed as a plug-and-play solution, GPCDR transforms single-domain GNN-based recommenders into effective cross-domain recommenders via the graph prompting approach. Specifically, to handle interest heterogeneity across distinct graph structures, we introduce a {multi-interest graph prompting module}. It employs a Mixture-of-Experts (MoE) mechanism to generate adaptive, user-specific graph prompts, enabling the backbone to capture diverse interest patterns. Furthermore, to address the rigidity of interest alignment, we design a {personalized domain similarity module}. Instead of enforcing uniform constraints, this module learns user-specific affinity scores to dynamically gate knowledge transfer, accommodating users with varying degrees of cross-domain consistency. Extensive experiments on four real-world datasets demonstrate that GPCDR not only significantly boosts the performance of various GNN backbones but also consistently outperforms state-of-the-art specialized CDR baselines. Our code is available at \url{https://github.com/liujinghong123/GPCDR}. \end{abstract}

Cross-Domain Sequential Recommendation (CDSR) mitigates data sparsity and cold-start issues by transferring user preferences across domains. Although Graph Neural Networks (GNNs) effectively capture collaborative signals, they struggle with the inherent structural sparsity present in individual domains. Specifically, users or items with few interactions become "isolated nodes". Traditional cross-domain methods fail to rescue these nodes because they rely on explicit, overlapping interaction topologies that are often missing, severely hindering message passing. In contrast, Large Language Models (LLMs) encode rich semantic knowledge yet lack the inherent ability to interpret collaborative graph structures. To address this limitation, we propose SGLP (Semantic-Guided Graph Structure Learning with LLM Structural Priors), a framework that leverages LLM knowledge as a structural prior to reconstruct sparse interaction graphs. Specifically, SGLP employs LLMs to generate Item Semantic Embeddings and User Global Semantic Profiles, establishing a robust semantic foundation. Distinct from traditional Graph Structure Learning (GSL) methods that rely solely on intrinsic node features, SGLP introduces a Semantic-Guided Graph Modeling module. This module applies differentiable structure learning to dynamically construct a task-adaptive hybrid graph, transforming unstructured semantic similarities into structured collaborative connections. To resolve distribution discrepancies between the reconstructed structural and semantic views, we enforce a Semantic-Behavioral Alignment constraint via contrastive learning. Furthermore, a Bi-Level Sequence Prediction strategy is introduced to alleviate negative transfer arising from domain shifts. Extensive experiments on three real-world Amazon datasets demonstrate that SGLP outperforms state-of-the-art baselines, validating the efficacy of integrating semantic priors into graph structure learning for cross-domain recommendation.

Cross-Domain Sequential Recommendation (CDSR) has emerged as a promising paradigm for alleviating data sparsity by transferring user preferences across domains. However, existing methods typically rely on sparse overlapping users or static LLM-derived embeddings, leading to two challenges: the Overlap Dilemma, where fragmented cross-domain connectivity impedes effective knowledge transfer, and the Semantic Gap, where identical items lack domain-specific contextual interpretations. To address these challenges, we propose \textbf{D}omain-\textbf{A}ware \textbf{H}ierarchical \textbf{S}emantic \textbf{R}epresentation (\textbf{DA-HSR}), a unified framework for learning transferable yet domain-sensitive representations. Specifically, we develop a Hierarchical Semantic Representation module that employs shared codebooks with domain-adaptive codeword refinements through Domain-adaptive Residual Quantization, thereby disentangling global static knowledge from layered dynamic semantics. We further design a Sequence-aware Routing Fusion module that treats the output of each codebook layer as an independent representation component and applies sequence-conditioned cross-attention to select the most relevant semantic granularity for each user context. To ensure robust cross-domain transfer, we introduce Cross-domain Sequence Encoding to retrieve informative global histories, together with Denoising Cross-domain Interest Transfer to suppress negative noise during preference alignment. Experiments demonstrate that DA-HSR significantly outperforms state-of-the-art baselines while effectively bridging both the semantic and data gaps in CDSR. Code is available at \url{https://github.com/zaixiaxiaowu/DA-HSR}.

Jointly modeling search and recommendation has emerged as a promising paradigm to capture diverse user preferences. However, the undifferentiated integration of cross-domain behaviors often induces negative transfer, driven by intrinsic intent ambiguity and extrinsic cross-domain semantic interference. To address this, we propose DiffuSAR, a diffusion-based framework designed to purify cross-domain behavioral sequences. First, we introduce a dual-branch intent disentanglement module to regularize domain-specific behaviors into discrete intent prototypes, which serve as explicit semantic anchors for the denoising process. Second, we devise a masked autoregressive denoising mechanism to efficiently suppress cross-domain semantic interference, leveraging preceding target-domain behaviors as guidance signals. Finally, we incorporate a trajectory loss to guide the reverse diffusion trajectory, progressively aligning cross-domain representations with target intents in a coarse-to-fine manner. Extensive experiments on two public datasets demonstrate that our approach consistently outperforms state-of-the-art baselines on both recommendation and search tasks, validating the efficacy of generative denoising in mitigating negative transfer.

Cross-Domain Recommendation (CDR) mitigates sparsity by transferring knowledge from source to target domains, yet its performance heavily depends on the quality of negative samples. Uniform random sampling is dominated by easy negatives that yield vanishing gradients, while hard negative sampling based on point-wise scores may suffer from false negatives and limited coverage of the global item structure. We propose SGNCDR, a progressive framework that unifies structure-aware negative selection and diffusion-based negative generation. SGNCDR selects diverse negatives from the observed item set by accounting for the global item-space structure, and constructs user-conditioned negative representations with a diffusion model using both domain-invariant and domain-specific user signals. With difficulty-controlled guidance, SGNCDR generates negatives at multiple hardness levels and integrates them into training via an easy-to-hard curriculum to refine preference modeling while maintaining stable optimization. Experiments on benchmark datasets show that SGNCDR consistently outperforms representative CDR baselines, and analyses confirm the effectiveness of each component. \textit{\href{https://github.com/fanydan/SGNCDR}{The codes are available here.}}

In online advertising, marketing interventions such as coupons introduce significant confounding bias into Click-Through Rate (CTR) prediction. Observed clicks reflect a mixture of users' intrinsic preferences and the uplift induced by these interventions. This causes conventional models to miscalibrate base CTRs, which distorts downstream ranking and billing decisions. Furthermore, marketing interventions often operate as multi-valued treatments with varying magnitudes, introducing additional complexity to CTR prediction. To address these issues, we propose the \textbf{Uni}fied \textbf{M}ulti-\textbf{V}alued \textbf{T}reatment Network (UniMVT). Specifically, UniMVT disentangles confounding factors from treatment-sensitive representations, enabling a full-space counterfactual inference module to jointly reconstruct the debiased base CTR and intensity-response curves. To handle the complexity of multi-valued treatments, UniMVT employs an auxiliary intensity estimation task to capture treatment propensities and devise a unit uplift objective that normalizes the intervention effect. This ensures comparable estimation across the continuous coupon-value spectrum. UniMVT simultaneously achieves debiased CTR prediction for accurate system calibration and precise uplift estimation for incentive allocation. Extensive experiments on synthetic and industrial datasets demonstrate UniMVT's superiority in both predictive accuracy and calibration. Furthermore, real-world A/B tests confirm that UniMVT significantly improves business metrics through more effective coupon distribution.

Generative modeling, particularly Decision Transformers, shows promise in auto-bidding strategies. However, it struggles with limited exploration in static offline datasets, vulnerability to dynamic environmental shifts, and a reliance on heuristic return-to-go (RTG) specifications that may yield suboptimal performance on unseen traffic. To address these challenges, we propose the Two-stage Uncertainty-aware Generative Auto-Bidding (TUGB) framework. TUGB integrates generative sequence modeling with Reinforcement Learning (RL). It employs a two-stage pipeline where a pre-trained Generative Bidding Model is fine-tuned via an ensemble-based Uncertainty-aware Environment Model. This approach quantifies epistemic uncertainty to penalize high-risk trajectories and evolves the policy from RTG-conditioning to direct bidding optimization. Extensive experiments demonstrate that TUGB significantly outperforms SOTA baselines. Notably, full-scale deployment on JD Search Ads achieved a 2.32\% increase in cost and 3.57\% in target cost, validating its industrial efficacy, while its successful adaptation to subsequent mechanism upgrades and JD Food Delivery Ads further validates TUGB's superior generalization capability.

Optimizing a single advertising campaign across heterogeneous channels is a central challenge in industrial autobidding. Auction mechanisms vary across channels in ranking rules (pure eCPM vs. UE-augmented scoring), pricing formats (first- vs. second-price), and bidding conventions (uniform vs. non-uniform), while advertisers impose shared campaign-level constraints. We propose HOB, which makes marginal cost (MC) computable and alignable across heterogeneous channels, especially for first-price auctions (FPA) with organic–paid coexistence, where existing bidding formulations do not yield a practical aligned MC form. At the global level, HOB derives channel-specific MC forms and coordinates disparate channels through a shared MC target. At the local level, HOB models free-win probability and winning-price uncertainty with a zero-inflated exponential distribution, yielding an efficient surplus-optimal bidding strategy for non-uniform first-price auctions. We show that any interior optimum satisfies MC equalization across channels. Experiments on a controlled offline benchmark, industrial log replay, and large-scale online A/B tests demonstrate that HOB consistently delivers significant performance gains. Deployed on a large-scale commercial DSP, HOB delivers a 3.0\% lift in GMV while maintaining return on advertising spend (ROAS) constraints.

Online advertising systems typically use a cascaded architecture to manage massive requests and candidate volumes, where the ranking stages allocate traffic based on eCPM (predicted CTR × Bid). With the increasing popularity of auto-bidding strategies, the discrepancy between the computationally sensitive retrieval stage and the ranking stages becomes increasingly critical, as the former cannot access precise, real-time bids for the vast ad corpus. This misalignment inevitably leads to revenue leakage for the platform and sub-optimal ROI for advertisers. To tackle this problem, we propose \textbf{B}idding-\textbf{A}ware \textbf{R}etrieval (\textbf{BAR}), a model-based retrieval framework that addresses multi-stage discrepancy by incorporating ad bid value into the retrieval scoring function. The core innovation is Bidding-Aware Modeling, utilizing bid signals through monotonicity-constrained learning and multi-task distillation to ensure economically coherent representations, while Asynchronous Near-Line Inference enables real-time updates to the embedding for market responsiveness. Furthermore, the Task-Attentive Refinement module selectively enhances feature interactions to disentangle user interest and commercial value signals. Extensive offline experiments and full-scale deployment across Alibaba's display advertising platform validated BAR's efficacy: 4.32\% platform revenue increase, with a 22.2 percentage-point gain in the Impression Improved Ratio for ads receiving positive operational adjustments.

Embedding-based retrieval (EBR) for large-scale e-commerce search faces three intertwined challenges: graded (non-binary) relevance where engagement signals are noisy and intent-varying while business relevance guidelines admit acceptable-but-not-exact matches, false negatives in hard sample mining, and unstable similarity score separability across relevance levels, the last of which complicates hybrid search score fusion and downstream ranking. We propose \emph{Mine and Refine}, a two-stage contrastive training framework that addresses all three. A lightweight LLM, fine-tuned with engagement-driven audit, serves as a guideline-aligned scalable labeler throughout training. Stage 1 establishes a robust global embedding space via label-aware supervised contrastive learning; Stage 2 mines hard samples, re-annotates them with the LLM labeler to mitigate spurious negatives, and refines the model through a multi-level extension of circle loss that enforces margin-controlled separation across relevance levels. Deployed in production e-commerce search across multiple product verticals, the approach delivers statistically significant lifts in user engagement and gross order value, and substantially improves retrieval and end-to-end relevance metrics.

Associating user search queries with the correct brand entity is critical for e-commerce product retrieval, yet remains challenging due to the brevity of queries (three to four words on average), their lack of grammatical structure, and a catalog of hundreds of thousands of distinct brands. We formulate this as a brand entity linking task and develop two complementary solutions deployed at scale: (1)~a cascaded pipeline that first detects brand mentions via sequence labeling and then disambiguates against a brand knowledge base, and (2)~a single-stage approach that frames linking as extreme multi-class classification, directly mapping queries to brand identifiers. Through extensive multilingual evaluation (11 languages) and a controlled online experiment, we demonstrate that the proposed methods substantially improve brand recall while maintaining high precision, leading to measurable gains in customer engagement.

E-commerce search often suffers from vocabulary mismatch between user queries and merchant-authored product titles, since short titles cannot fully cover diverse user expressions or visual product attributes. Although Doc2Query alleviates this issue by generating pseudo-queries for document expansion, traditional methods are text-only and not optimized for e-commerce business objectives. As a result, they may produce semantically plausible but commercially ineffective expansions and miss key attributes present in product images. To this end, we propose E-commerce Search-Aligned Multimodal Doc2Query (SAM-D2Q), a business-aligned multimodal document expansion framework for e-commerce search under Boolean retrieval constraints. SAM-D2Q consists of three stages: (1) task-adapted multimodal supervised fine-tuning to enhance vision-language understanding of product titles, images, and user queries; (2) multimodal data augmentation to improve perception of key visual attributes and expansion coverage; and (3) reinforcement-learning-based preference alignment toward search business objectives, encouraging the model to generate pseudo-queries that better match user intent and commercial value. Offline experiments show that SAM-D2Q substantially improves retrieval performance over traditional Doc2Query methods. Deployed in the AliExpress production search system, SAM-D2Q improves online business metrics, increasing GMV by +3.38% and Pay Count by +2.27%.

The rapid rise of large language models has shifted user search behavior from discrete keywords to natural-language, multi-constraint queries—a shift existing e-commerce search architectures struggle to accommodate. Users face a dilemma: precise natural-language queries often trigger zero-result scenarios, while forced simplification yields noisy, generic results that overwhelm decision-making. To address this, we propose LEAPS (LLM-Empowered Adaptive Plugin in Taobao AI Search), which upgrades traditional search pipelines via a "Broaden-and-Refine" paradigm by attaching plugins at both ends. (1) Upstream, a Query Expander generates adaptive, complementary query combinations to maximize the candidate set, trained via a three-stage strategy of inverse data augmentation, posterior-knowledge supervised fine-tuning, and diversity-aware reinforcement learning. (2) Downstream, a Relevance Verifier filters noise by synthesizing multi-source signals (e.g., OCR text, reviews) with chain-of-thought reasoning. Extensive offline experiments and online A/B testing show that LEAPS significantly enhances the conversational shopping experience, while its non-intrusive architecture preserves established short-text retrieval performance and enables low-cost integration with diverse back-ends. Fully deployed on Taobao AI Search since August 2025, LEAPS serves hundreds of millions of users monthly.

Search relevance modeling is a core task in e-commerce search systems, assessing how well a user query matches candidate products. Rather than relying on a single holistic matching signal, relevance judgment often requires structured reasoning over query understanding, product understanding, and facet-level matching. With large language models (LLMs), this process is increasingly formulated as chain-of-thought (CoT) reasoning and optimized with reinforcement learning (RL). However, existing RL methods mainly rely on outcome-level rewards and treat the entire reasoning chain as a single optimization unit. This makes it difficult to distinguish faulty reasoning steps from correct intermediate ones, leading to misaligned credit assignment. Although process-reward methods provide denser supervision, they often treat reasoning steps independently and ignore dependency-driven error propagation, making responsibility attribution difficult and limiting the optimization of structured relevance reasoning. We propose Graph-GRPO, a graph-structured extension of GRPO for multi-component relevance reasoning. Graph-GRPO constructs a relevance reasoning dependency graph, where CoT steps are modeled as nodes and their logical dependencies as edges. It propagates outcome-level rewards over the graph to derive step-level credit signals, enabling more accurate fine-grained credit assignment. We further introduce a main-loss-driven controller that adaptively adjusts edge-wise credit-propagation coefficients. Together with CoT random masking for supervised policy initialization and graph-node-based multi-head distillation, we build a trainable and deployable framework for generative relevance modeling. Extensive offline evaluations and online A/B tests on a leading e-commerce platform demonstrate that the Graph-GRPO-based framework improves relevance classification metrics and key engagement metrics. The resulting lightweight model has been deployed in the JD e-commerce search system, serving hundreds of millions of users.

Generative retrieval offers a promising alternative by unifying the multi-stage retrieval process into a single end-to-end model. However, its practical adoption in industrial e-commerce search remains challenging, given the massive and dynamic product catalogs, strict latency requirements, and the need to align retrieval with downstream ranking goals. In this work, we propose a retrieval framework tailored for real-world recall scenarios, positioning generative retrieval as a recall-stage supplement rather than an end-to-end replacement. Our method, CQ-SID (Category-and-Query constrained Semantic ID), employs category-aware and query-item contrastive learning along with Residual Quantized VAEs to encode items into hierarchical semantic cluster identifiers, significantly improving click and exposure metrics and reducing beam search size and inference time. Additionally, we develop EG-GRPO (Expert-Guided Group Relative Policy Optimization), a reinforcement learning approach that aligns generative recall with downstream ranking under sparse rewards by injecting ground-truth samples to stabilize training. Offline experiments on TmallAPP search logs show that CQ-SID achieves up to 26.76% and 11.11% relative gains in semantic and personalized click hitrate over RQ-VAE baselines. EG-GRPO further improves multi-objective performance. Online A/B tests confirm gains in GMV (+1.15%) and UCTCVR (+0.40%). The generative recall channel now contributes substantially in production, accounting for 50.25% of exposures, 58.96% of clicks, and 72.63% of purchases, demonstrating a viable path for deploying generative retrieval in real-world e-commerce systems.

Embedding-based retrieval (EBR) is a prevailing method in modern e-commerce search, retrieving relevant items from billions of candidates in response to users' queries. While recent methods often fine-tune large language models (LLMs) for representation learning, they typically lack robust mechanisms for handling complex and implicit semantics. To address this, Retrieval-GRPO (R-GRPO) was recently proposed as a multi-objective Reinforcement Learning (RL) framework for dense retrieval. It computes rewards using top-$K$ items dynamically retrieved per query. While this approach partially mitigates retrieval challenges, the limited candidate pool sampled per batch often introduces noisy samples into the top-$K$ results. Moreover, the LLM used for computing relevance rewards in R-GRPO is trained with an RL framework similar to R-GRPO itself. This inherent similarity makes it difficult to provide unbiased and impartial assessment results. To tackle these issues, we propose Supervised Retrieval-GRPO with Semantic Identifiers (SSR-GRPO). Specifically, our method first proposes a dual-perspective framework for relevance assessment. It leverages both Semantic Identifiers (SIDs) produced by quantization learning and dense representation vectors to generate more unbiased relevance scores. Furthermore, leveraging the hierarchical similarity relationships of the generated SIDs, we mine a set of hard negative samples that serve two purposes: (1) to design a masking function integrated into R-GRPO, effectively filtering intra-group noisy samples; and (2) to construct a Retrieval-DPO task composed of positive and negative sample pairs, enabling the model to capture fine-grained semantic distinctions from a pair-wise perspective. By integrating these optimization strategies, we propose SSR-GRPO. Extensive offline and online experiments validate SSR-GRPO's effectiveness, and it has been deployed on a large-scale e-commerce platform.

Semantic search with large language models (LLMs) enables retrieval by meaning rather than keyword overlap, but scaling it requires major inference efficiency advances. We present LinkedIn’s LLM-based semantic search framework for AI Job Search and AI People Search, combining an LLM relevance judge, embedding-based retrieval, and a compact Small Language Model trained via multi-teacher distillation to jointly optimize relevance and engagement. A prefill-oriented inference architecture co-designed with model pruning, context compression, and text–embedding hybrid interactions boosts ranking throughput by over 75× under fixed latency constraint while preserving near–teacher-level NDCG, enabling one of the first production LLM-based ranking systems with efficiency comparable to traditional approaches and delivering significant gains in quality and user engagement.

Speech toxicity detection is increasingly important in real-time voice-based interactions, where harmful speech can affect listeners before an utterance is fully observed. Existing approaches often treat speech as either transcribed text or a single audio embedding. However, toxic expressions can be lexically diverse, while vocal delivery may contain recurring acoustic-prosodic cues such as intensity, stress, rhythm, voice quality, or sarcastic tone. Motivated by this, we propose a latent query-guided acoustic-prosodic retrieval framework that learns class-specific latent queries to retrieve prediction-relevant local cues from speech tokens without relying on ASR transcripts or manually written prompts. Experiments on the combined English DeToxy+MuTox datasets show that our framework outperforms single-encoder, fusion, and retrieval baselines and improves early detection under prefix cropping. Analyses further show that the learned queries attend to temporally diverse regions and capture acoustic-prosodic differences associated with toxic and non-toxic speech. Latency measurements also indicate a lower-latency alternative to ASR-based text pipelines for real-time toxicity detection.

Memes, widely shared on the internet as a form of cultural expression, present a challenge in identifying harmful content due to their compact format and subtle meanings. Current detection methods employ pre-trained models to extract image and text features from memes, subsequently combining these features through deep neural networks for classification. However, these approaches often overlook the underlying meanings embedded in memes, focusing solely on shallow semantics and surface-level harmful signals. This makes it difficult to capture the cultural connotations and subtle expressions inherent in memes. Moreover, they lack explanations for their judgments. To address these challenges, we propose a novel framework for explainable harmful meme detection. This framework combines the advantages of Multimodal Large Language Models (MLLMs) and fine-grained heterogeneous semantic graphs. By leveraging multi-hop reasoning, we utilize MLLMs' extensive knowledge to provide comprehensive meme analyses, addressing meme literacy gaps. We extract key objects, relationships, and attributes from images to produce detailed scene graphs, which support a deeper understanding of the meme. The fine-grained heterogeneous semantic graphs filter and integrate multiple information sources to uncover hidden harmful elements. Additionally, through multi-hop reasoning and scene graphs, our approach generates reliable explanations via MLLMs. Experimental results demonstrate that our method outperforms existing state-of-the-art approaches on widely used datasets.

Multimodal Emotion Recognition in Conversation (MERC) aims to identify the emotion of each speaker in every utterance by leveraging multimodal conversational signals. Existing methods typically employ graph structures to model contextual dialogue interactions; however, they often neglect fine-grained affective correspondences and implicitly assume equal reliability across modalities, making multimodal representations susceptible to modality-specific noise and semantic interference. To address these limitations, we propose SFAC, a semantic-guided fine alignment and asymmetric contrast framework with alignment-aware conversational graph modeling for MERC. Specifically, an optimal-transport-based fine-grained alignment objective captures flexible many-to-many correspondences among textual tokens, acoustic frames, and visual frames before conversational graph modeling. In addition, a semantic-guided asymmetric contrastive mechanism employs stop-gradient textual representations as semantic anchors to regularize acoustic and visual modalities after graph-based contextual reasoning, thereby preserving complementary multimodal information while mitigating interference from noisy auxiliary signals. Finally, the refined multimodal representations are fed into a decoder that formulates emotion recognition as an autoregressive text generation task. Extensive experiments on the IEMOCAP and MELD benchmarks demonstrate that SFAC consistently outperforms state-of-the-art methods, highlighting the effectiveness of semantic-guided fine alignment and asymmetric contrastive learning. Our code is available at \url{https://github.com/cgao-comp/SFAC}.

Inferring human sentiment through the combined processing of language, audio, and vision signals is the core focus of multimodal sentiment analysis. Its effectiveness critically depends on high-quality multimodal representations that capture both modality-invariant and modality-specific features, while properly modeling the unequal contributions of different modalities. To address these challenges, we propose DGR-Net, a framework that jointly performs feature decomposition, hybrid dynamic gating, and complementary retrieval in a synergistic manner. Unlike prior purely work, our hybrid dynamic gating (HDG) module combines semantic attention with energy-based uncertainty for sample-level modality weighting, and our modality-specific retrieval (MSR) module is designed to selectively extract key cues from modality-specific features, thereby boosting the final prediction performance. Extensive experiments on MOSI and MOSEI demonstrate that DGR-Net achieves highly competitive performance and validate the contribution of each proposed module. Code is available at \url{https://github.com/cgao-comp/DGR-Net}.

Multimodal learning combines heterogeneous signals, such as text, audio, and vision, to improve prediction. However, adding modalities does not always help, because modalities may contain useful complementary cues together with noisy, redundant, or contextually unreliable components that interfere with full-modality learning. In this work, we propose Cross-Modal Value-Gated Modality Refiner (VGMR), a pre-fusion refinement module that estimates context-conditioned value evidence before modality features enter the backbone. VGMR models summary-level cross-modal agreement and discrepancy, estimates global-level and channel-level value signals, and uses them with raw and projected features to generate fine-grained gates over the original modality features. Different from conventional attention or generic gating, VGMR treats value as signed, context-conditioned impact evidence rather than a direct preservation probability, allowing the gate generator to learn whether each component should be enhanced, retained, or suppressed. We evaluate VGMR on MOSI and MOSEI with a Transformer backbone, and on MOSI, MOSEI, UCF101, AVE, and CREMA-D with a ResNet+Concat backbone. On MOSEI with the Transformer backbone, VGMR achieves 0.8446 Accuracy and 0.8029 Macro-F1, outperforming the strongest baselines by 1.09 and 0.89 percentage points, respectively. Ablation, controlled comparison, modality removal, noise robustness, optimization, and qualitative analyses suggest that VGMR can reduce unreliable modality responses before fusion and improve robustness under noisy multimodal inputs.

Multimodal learning often suffers from modality imbalance, leading to an optimization dilemma. A key reason is that different modalities exhibit asynchronous learning dynamics during joint training. A representative line of existing methods seeks to rebalance multimodal learning through cross-modal alignment. However, such alignment is usually imposed in a static, symmetric, and fully coupled manner, making it prone to unreliable cross-modal interactions. To address this issue, we argue that cross-modal alignment under modality imbalance should be space-aware and status-aware, reducing feature-space discrepancy while selectively aligning transferable components according to modality learning status. Based on this principle, we propose an \textbf{A}daptive \textbf{B}idirectional \textbf{A}lignment (ABA) method. ABA first constructs class-conditional statistical projections to establish more robust semantic correspondences between heterogeneous modality spaces. It then employs feature-level partial optimal transport to selectively align reliable and transferable feature components. Furthermore, we introduce a dynamic bidirectional regulation mechanism to adaptively adjust the direction and strength of cross-modal knowledge transfer according to modality learning status, enabling more effective cross-modal collaboration. Empirical experiments on widely used datasets reveal the superiority of our method through comparison with various state-of-the-art (SOTA) multimodal learning baselines. The source code is available at \href{https://github.com/yilnr/ABA}{our repository

Imbalanced multimodal learning remains a critical challenge, as models often over-rely on dominant modality while under optimizing non-dominant one, ultimately limiting overall performance. Essentially, we observe that such over-reliance is closely associated with discrepancies in feature aggregation capability across modalities. More concretely, non-dominant modality lags in forming compact and stable representations, whereas dominant modality exhibits superior aggregation capability. To address this issue, we propose a simple yet effective approach called \underline{D}ynamic \underline{S}ample \underline{R}etrieval for \underline{E}fficient \underline{F}eature \underline{A}ggregation (\textbf{DSR-EFA}). Specifically, we first introduce a retrieval-driven near-neighbor sample augmentation strategy to build a candidate pool for each modality. Based on these pools, we propose a modality-aware dynamic ensemble scheme that allocates retrieval budget according to modality importance and constructs additional informative samples for classification. Consequently, DSR-EFA effectively reduces feature aggregation disparities and improves multimodal learning performance. Extensive experiments on widely used datasets verify the effectiveness of DSR-EFA.

Multimodal Emotion Recognition (MER) leverages diverse modalities such as text, vision, and audio to infer emotional states, yet suffers from prohibitively high annotation costs. As a promising solution, Active Learning (AL) strategically selects the most valuable data for annotation, enabling competitive model performance with substantially reduced labeling effort. To address this issue, this study proposes MALCC, a novel multimodal active learning framework that leverages emotion-aware cross-modal contrastive learning for emotion recognition. Specifically, to achieve task-specific alignment across modalities, we first design an emotion-aware cross-modal contrastive learning module, which embeds multiple modalities into a unified emotion-aligned representation space. Within this space, we propose a comprehensive reliability-aware query strategy that prioritizes cross-modal semantic consistency and modality contribution balance as two key reliability criteria, while incorporating informativeness and representativeness as complementary utility criteria to select more valuable instances for annotation. By leveraging this strategy, we then design a novel multimodal active learning algorithm tailored for emotion recognition. We further provide theoretical analysis to justify our design and conduct extensive experiments on three widely used MER datasets, demonstrating superior performance over state-of-the-art (SOTA) methods. Our code and supplementary materials are available at https://github.com/Atlantis-esh/malcc0425.

Complex temporal knowledge graph question answering (TKGQA) requires composing multiple time-dependent facts under constraints such as validity intervals, temporal ordering, overlap, and set-style operations. Recent LLM-based approaches improve temporal reasoning through retrieval, decomposition, or planning, but the temporal constraints produced during intermediate reasoning are often left implicit: retrieved evidence may be temporally invalid, intermediate answers may be propagated without verification, and deterministic temporal operations may still be handled by free-form generation. We propose TRACE, a temporal-constraint-aware recursive framework that traces temporal constraints across retrieval, answer verification, constraint propagation, and execution. TRACE grounds temporal anchors and operators from sub-questions and intermediate answers, performs operator-aware temporal retrieval, verifies child-derived answers using grounding, bridge evidence, and operator consistency, and invokes symbolic temporal execution when root-level temporal operations are deterministically resolvable. Experiments on MultiTQ and TimelineKGQA show that TRACE improves multi-step and operator-sensitive temporal reasoning under the evaluated settings. Detailed analyses further show that verified answer propagation reduces error accumulation, while symbolic temporal execution is particularly effective for compositional temporal relations such as intersection and union. These results demonstrate that complex TKGQA benefits not only from retrieving or decomposing temporal questions, but also from explicitly tracing, verifying, and executing temporal constraints throughout the reasoning process.

Temporal question answering over temporal knowledge graphs (TKGQA) requires retrieving evidence that is not only semantically relevant but also temporally valid. However, existing retrieval-augmented generation (RAG) approaches primarily rely on semantic retrieval and often fail to enforce temporal constraints during retrieval, leading to temporally inconsistent evidence, especially for questions involving multiple events and implicit temporal dependencies. In this paper, we propose a time-constrained retrieval-augmented question answering framework for temporal knowledge graphs. Given a temporal question, our method decomposes it into an ordered sequence of structured event queries, each optionally associated with a temporal constraint and an event-time variable. Retrieval is conducted in a stepwise manner, where temporal constraints are enforced throughout the retrieval process via hard temporal filtering followed by semantic--temporal fusion re-ranking. Semantic relevance is modeled jointly from entity-, relation-, and sentence-level representations, while temporal relevance is incorporated as an additional ranking signal. We evaluate our approach on two temporal question answering benchmarks, MultiTQ and CronQuestions. CronQuestions involves year-level interval facts, whereas MultiTQ uses day-level time-stamped facts and includes questions with mixed temporal granularities and compositional temporal constraints. Experimental results show that our method achieves the best overall Hits@1 among the compared methods on both benchmarks. Ablation studies on MultiTQ further validate the effectiveness of structured temporal decomposition and time-constrained retrieval.

Link prediction on temporal knowledge graphs (TKGs), a.k.a. TKG link prediction, aims to predict missing facts from historical facts. Existing rule-based methods mainly rely on sequential temporal rules that propagate variables along chronological paths. However, such path-shaped rules cannot naturally capture a common prediction pattern in which a historical source-target interaction first identifies a candidate target, while subsequent source-side events provide contextual support before the query time. In this paper, we propose a new class of star-shaped temporal rules (SSRs) for TKG link prediction and a rule-based framework called PLSSR that discovers and applies SSR rules for link prediction. To discover SSRs, PLSSR employs a top-?? pattern growth strategy over localized contextual relations. During inference, SSR rules are aggregated through Noisy-OR scoring, while sparse candidate sets are supplemented using subject-history and relation-level target priors. Experimental results demonstrate that PLSSR outperforms current embedding-based, rule-based, and LLM-assisted baselines with average relative MRR gains of 78.3%, 66.4%, and 57.9%, respectively, while preserving explicit, interpretable reasoning traces.

Extrapolative temporal knowledge graph reasoning (TKGR) predicts future facts from historical snapshots. Most existing methods train once on an early prefix of the timeline and then use a frozen model for all future timestamps. We argue that this fixed-prefix protocol is misaligned with extrapolation. It learns from a static prefix, whereas the target stream is non-stationary: new entities and facts emerge, temporal dependencies shift across regimes, and recurring historical signals must be refreshed online. As a result, models trained only on early snapshots become outdated and degrade over long horizons. We address this mismatch by formulating extrapolative TKGR as continual learning over streaming snapshots. Under this view, effective extrapolation must jointly handle \emph{current dynamics}, \emph{stable knowledge}, and \emph{recurring historical evidence}. Based on these requirements, we propose \textbf{Hi}story-enhanced \textbf{T}wo-\textbf{S}tep \textbf{C}ontinual \textbf{L}earning (\textbf{HiTS-CL}), a backbone-agnostic continual learning framework for extrapolative TKGR. HiTS-CL tracks current dynamics via continual fine-tuning, preserves stable knowledge via multi-teacher adaptive distillation, and retains recurring historical evidence via a selective memory of recent and frequent facts. We integrate HiTS-CL into five representative TKGR backbones and evaluate it on four benchmark datasets. HiTS-CL consistently improves extrapolation accuracy, reduces long-horizon degradation, and outperforms strong continual-learning baselines, including a recent method for temporal knowledge graphs. Source code and data are available at \url{https://github.com/liuyansong98/HiTS-CL}.

Event knowledge graphs (EKGs) must evolve continuously as new events emerge, while preserving temporal integrity and structural consistency. However, existing triple-completion and embedding approaches assume closed-world settings and lack mechanisms for open-world alignment and interpretable updates. Furthermore, real-world streaming data often suffers from out-of-order arrivals and information fragmentation, which exacerbates update conflicts and causal inversion in continuous evolution. We define the task of Continuous Event Knowledge Graph (C-EKG) Updating and introduce \textbf{HGNN-EAGE}, a unified framework that addresses these challenges through adaptive stream scheduling and hypergraph neural alignment. Specifically, we propose an adaptive dependency-aware sorting strategy that dynamically prioritizes information-rich and temporally antecedent events to initialize the topological basis. For event alignment, HGNN-EAGE models events as hyperedges and learns numerically stable embeddings via a degree-regularized hypergraph neural network. It decouples semantic, structural, temporal, spatial, type-level, salient-entity, and consistency cues into orthogonal evidence dimensions, driving a hierarchical decision process that maps interpretable alignment outcomes to atomic, version-controlled graph operations (insert, supplement, link-subevent). Extensive experiments on EventKG and DuEE-fin benchmarks demonstrate that our framework significantly outperforms strong baselines in retrieval and classification accuracy. Moreover, the proposed streaming strategy substantially mitigates topological redundancy under high-load out-of-order data streams while maintaining near-linear scalability.

Knowledge graphs (KGs) rely on structured representations of entities and relations to enable effective information access and integration. Building and populating KGs from semi-structured sources such as tables remains a challenging task, particularly due to the need to adhere to strict ontology constraints on valid entity types and relations. Large Language Models (LLMs) provide strong capabilities for understanding and generalising from text, but ensuring that their outputs conform to a target KG schema without explicitly providing the full ontology remains an open problem. In this paper, we propose WikiTabLLM, a few-shot, example-driven framework for ontology-compliant knowledge graph population. Instead of explicitly specifying ontology constraints, WikiTabLLM leverages partially linked Wikipedia tables to implicitly induce valid typing and relation patterns. Each table is partitioned into (i) rows containing entities already linked to the Wikidata knowledge graph, which serve as in-context supervision signals, and (ii) rows with unlinked entities, for which we aim to generate RDF triples for KG population. WikiTabLLM learns to infer valid properties and relations by conditioning on the structured patterns exhibited in existing triples, effectively enforcing ontology compliance through examples. We evaluate WikiTabLLM on a newly constructed dataset of Wikipedia tables and compare against Semantic Table Interpretation, relation extraction baselines, and zero-shot LLM prompting. Results show that our approach produces more accurate and schema-consistent triples (up to a 3x improvement in macro F1), demonstrating the effectiveness of leveraging in-context examples to guide ontology-constrained knowledge graph population.

Constructing formal ontologies from domain documents requires simultaneously enforcing corpus grounding, vocabulary consistency, axiom-level expressivity, and end-to-end provenance, a combination no existing automatic system delivers. We present a seven-stage graph-grounded pipeline that converts domain documents into a complete, auditable Web Ontology Language (OWL) Terminological Box (TBox) without any unconstrained generation step. Documents are first encoded as Unified Discourse-Hypergraphs (UDH) capturing entity participation and discourse dependencies; subsequent stages transform this graph evidence into a class hierarchy, typed object and datatype properties, and restriction axioms, with Large Language Model (LLM) usage restricted to narrow, graph-grounded mediation tasks. A paired Assertional Box (ABox) population procedure grounds named individuals in the induced TBox, enabling SPARQL-based functional evaluation. Every emitted term carries a full decision chain from raw source passages through each pipeline stage, making the TBox directly auditable and suitable for targeted human refinement. Evaluated on the life insurance domain using two established benchmarks and a new 100-contract corpus spanning ten product types, our pipeline achieves strong results across all evaluation dimensions, outperforming direct and multi-agent LLM baselines on competency-question (CQ) coverage (0.85 vs. 0.63 and 0.62 on one term-life contract; 0.77 vs. 0.40 and 0.44 on another contract), while also attaining high keyphrase coverage comparable to a manually-constructed reference ontology and strong performance on structured gap-and-overlap reasoning, all without any manual TBox engineering. Ontology growth analysis provides evidence consistent with vocabulary saturation at scale, demonstrating that the pipeline produces stable, reusable domain representations from large document corpora.

Modern LLM-based applications increasingly rely on external memory for storing prior interactions, user-provided files, enterprise records, and domain-specific documents that can be reused in later interactions. However, as memory grows, current infrastructures provide limited support for understanding what the memory contains, how it is organized, and where retrieved evidence comes from. Vector indexes expose memory mainly through flat top-k similarity search, Knowledge graphs often require predefined schemas and fine-grained conversion of raw information into entities and relations which is difficult to navigate, meanwhile existing taxonomy construction methods are commonly static or one-shot. These limitations make current approaches difficult to maintain for continuously growing LLM memory. We present TaxoEvolve, an incremental framework that organizes memory items into a self-evolving taxonomy. It routes items to existing nodes, creates classes for emerging concepts, and refines overloaded nodes using a distribution-based split score that favors coherent groups with enough support. The taxonomy exposes an inspectable, editable memory structure, allowing humans or LLM-based agents to browse, search selected regions, and trace evidence through class paths and history tags. On TaxoBench, TaxoEvolve achieves stronger organization quality than clustering-based and LLM-based baselines, while some one-shot LLM baselines cannot process the full input due to context-window limits. Across scientific, medical, and financial retrieval tasks, taxonomy-guided search improves Recall@20, reaching 0.918 with hybrid retrieval and 0.923 with tag-aware retrieval.

Modern information and knowledge systems increasingly rely on frozen neural content scorers to rank, filter, recommend, moderate, and curate image, audio, and multimedia content. For trustworthy deployment, operators need attribution tools that explain not only salient input regions, but which structured signal components can locally restore or strengthen the scalar scores consumed by downstream ranking, filtering, and human-review workflows. We propose Adaptive-Path Wavelet Interpretation Maps (AP-WIM), a reference-input-free attribution method that computes nonnegative ascent-sensitivity maps in an invertible wavelet coefficient space. AP-WIM starts from the observed sample, advances along the local gradient field under an amplitude-proportional budget, and accumulates score-increasing sensitivity for input-aligned frequency-structured heatmaps. We characterize the origin-neighborhood risk of zero-baseline straight-line paths and, under local smoothness, derive a conditional score-improvement lower bound; the sign update is first-order optimal within its prescribed coordinate-wise budget box. On ESC-50 audio classifiers, paired-bootstrap tests show significant insertion gains under community-standard perturbation protocols, and a five-fold attribution stability check confirms that the insertion advantage is not unique to one audio fold. On a scorer-level Tiny-ImageNet query-gallery ranking audit, AP-WIM achieves the highest mean insertion AUC across target-score, Recall@1/5, MRR, and nDCG@10 endpoints, and significantly improves relevant-item score restoration over the strongest non-AP baseline. Together, the ranking audit, multimodal attribution benchmarks, runtime measurements, and robustness checks position AP-WIM as a resource-aware, frequency-structured explainability interface for frozen neural content-scoring pipelines. Code is available at: https://github.com/Davidjinzb/AP-WIM.

Real-world decision-making in public health and social science can greatly benefit from predictive models, yet translating predictions into effective interventions requires explaining the model behavior. While Graph Neural Networks (GNNs) are well-suited for modeling relational data, existing explanation methods largely operate at the node level and fall short of supporting actionable, network-level intervention design. Existing counterfactual GNN explainers, such as CF-GNNExplainer and CF$^2$, rely on continuous mask optimization over features and edges, which implicitly assume feasible edge manipulation, may allocate effort to immutable or non-actionable attributes, and incur substantial computational overhead. Further, the method of arriving at the explanation itself is difficult to explain to a domain specialist who is not an AI expert. Can simple methods generate good explanations? To explore this, we reframe counterfactual explanation as an intervention design problem. At the local level, we generate counterfactuals via a greedy search that directly identifies minimal, actionable changes to node features and neighbor-level conditions. We derive conditions under which the greedy search provides guarantees, and empirically show that these conditions are approximately met. These counterfactuals are converted into interpretable rules suitable for real-world intervention. At the network level, we formulate intervention selection as a Disjunctive Normal Form (DNF) coverage problem under a budget constraint, which is nondecreasing and approximately submodular, enabling a greedy algorithm with theoretical guarantees. Experiments on synthetic graphs and real-world suicide risk networks demonstrate that our approach produces scalable, cost-effective intervention strategies with significantly improved efficiency over mask-based counterfactual methods.

Graph Neural Networks (GNNs) are widely used for learning from relational data, but they can also propagate and amplify biases associated with sensitive attributes. Recent fair GNNs have moved from single-level debiasing toward multi-level intervention, but the strength of these interventions is still commonly governed by fixed coefficients, task-specific modules, or manually selected objectives. We propose Adaptive Hierarchical Fairness Intervention (AHFI), a fair GNN framework that performs adaptive constraint control at the input, representation, and output levels. AHFI monitors group disparity at each level with a geometric fairness proxy and assigns level-specific fairness pressure through learnable dual variables. To stabilize optimization, AHFI uses scale-invariant constraint violations, adaptive moving targets, and a finite-sample floor that prevents the controller from over-penalizing noise-level disparity. Experiments on six standard fair node classification benchmarks show that AHFI provides favorable fairness-utility trade-offs compared with representative fair GNN baselines across common message-passing backbones. Ablation studies show that the multi-stage design and the adaptive controller both affect the final fairness-utility trade-off. These results suggest that stage-wise adaptive control is an effective strategy for fairness-aware graph learning. Codes can be found at https://github.com/Dr-HZH/AHFI.

Graph anomaly detection (GAD) identifies nodes deviating from normal graph patterns and is crucial in high-stakes applications such as fraud and social risk detection. However, existing GAD methods mainly prioritize accuracy while overlooking fairness. Sensitive information can be encoded in node attributes, homophilous links, and graph structures, and message passing may entangle these signals with anomaly evidence, causing group-imbalanced scores. Existing fair GAD methods are often detector-specific, limiting reuse and intensifying the fairness-utility trade-off. To address these issues, we propose Multi-Disentangled processing for Fair graph Anomaly Detection MDFairAD, built upon a Multi-Disentangled Fair Processor MDFP that serves as an upstream fair representation module jointly optimizable with off-the-shelf anomaly detectors. It performs multi-view disentanglement with causal- and representation-level debiasing to reduce attribute and structural leakage, and then nonlinearly fuses the disentangled branches to preserve cross-view anomaly evidence while suppressing sensitive information. The refined representations are further fine-tuned with anomaly detectors through a detection head. Experiments on real-world datasets show that MDFairAD achieves competitive detection performance while substantially improving fairness over existing baselines.

We propose the notion of Persistent-Bias-Aware (PBA) fairness and an approach to formally quantify locally-persistent bias. PBA fairness captures the bias in a model's predictions towards a certain outcome for a given input, compared to its counterfactual (where the sensitive attribute of the original input is flipped), in an entire region. As such, PBA fairness allows us to isolate the effect of the sensitive attribute, enabling formal identification of fairness disparities that are consistently present in the model behavior within a neighborhood. We evaluate our proposed approach under both fairness-agnostic and fairness-aware training methods and compare it to well-established fairness metrics on three benchmark datasets: Adult, COMPAS, and German Credit. The results demonstrate that our proposed approach identifies fairness disparities overlooked by other approaches and offers additional insights into model behavior. To address this challenge, we further propose a debiasing scheme by penalizing persistent bias. The results show that the models obtained under such debiasing scheme are consistently less prone to locally-persistent bias compared to the state of the art.

Tiered service platforms in mobility, delivery, and cloud computing accelerate premium requests, and aggressive prioritization can silently degrade the experience of regular users. We study premium admission as an auditable admission-control problem over streaming telemetry: given per-region arrival rates, service-time moments, and a declared tolerance $\delta$ on the mean delay inflation of regular users, how many premium requests may be prioritized? Building on a two-class M/G/1 model with probabilistic gap insertion and head-of-line promotion, we derive a closed-form fairness certificate that maps three telemetry statistics and the platform's prioritization parameters to an admissible premium-to-regular ratio, prove its monotonicity, and show that it is a conservative surrogate of the exact nonpreemptive-priority formula whenever promoted requests wait at most one mean service time. The certificate composes across heterogeneous regions through a per-region cap layer and a shared-slack pooling rule with explicitly declared per-region budgets, and it acts as a state-dependent gate for online controllers. We evaluate FairRank on trace-driven replays of NYC TLC high-volume for-hire vehicle records from three weeks of 2024, comparing nine admission policies, including an established utilization-threshold admission-control baseline, across three premium-candidate definitions and sweeps over tolerance, calibration, scale, and time. Certificate-gated admission reduces mean regular-user delay inflation about five-fold relative to unconstrained prioritization at the evaluated operating points, and the telemetry-updated adaptive variant attains the lowest inflation of all tested policies. Because the certificate is model-conditional, we also isolate which assumptions matter: a synthetic study shows that heavy-tailed service times leave the certificate intact while bursty arrivals roughly double realized inflation, and a diagnostic on the real trace explains the remaining gap and gives operators a conservative parameterization that meets the nominal tolerance in every zone.

Automated Machine Learning (AutoML) systems are increasingly deployed in settings where data arrives continuously and evolves over time. In such streaming environments, fairness presents a distinct challenge: data distributions change, sensitive-group proportions fluctuate, models are updated incrementally, and fairness measurements become noisy and time-varying. Existing fairness-aware AutoML approaches are designed for static data and enforce fixed fairness thresholds, while streaming AutoML systems largely neglect fairness. To address this gap, we introduce FairStream, a fairness-aware AutoML framework for data streams that treats fairness as a dynamic control signal. FairStream introduces a streaming evaluation framework that jointly models predictive performance and group-based fairness through signals capturing temporal variation, estimation uncertainty, and both performance and fairness drift. These signals drive an adaptive fairness control mechanism that maintains a time-varying feasibility envelope and accumulated fairness debt to enforce consistent fairness over time. We formulate model selection as a sequence of online constrained optimization problems and solve it using constrained Bayesian optimization, enabling principled exploration under evolving fairness constraints. Experiments on diverse streaming benchmarks show that FairStream achieves strong predictive performance while substantially reducing fairness violations compared to fairness-agnostic and fixed-threshold fairness-aware baselines.

\begin{abstract} Data valuation is a fundamental problem in emerging data markets, where the value of data is inherently shaped by both the market context and the internal structure of the dataset. However, existing approaches typically capture only one side of this interplay. Classical Shapley-based methods treat data value as an intrinsic property of a fixed dataset, leading to inconsistent valuations across participants, while recent approaches that rely on global data distributions improve consistency at the cost of overlooking structural redundancy and incurring substantial computational overhead. In this work, we argue that data valuation should be understood as a context–structure coupling problem, where value arises from the interaction between global context and local data structure. Based on this perspective, we propose CoShapley (\textbf{Co}ntextual-Structural \textbf{Shapley}), a hierarchical valuation framework that explicitly models this coupling, producing valuations that are both stable across varying market contexts and sensitive to redundancy and correlation within the data. To make this formulation scalable, we introduce a signed decomposition that removes pairwise coupling in Shapley estimation, together with a hybrid two-stage sampling strategy that enables batch updates of all data points from a single model evaluation. Extensive experiments on multiple real-world datasets demonstrate that CoShapley achieves superior valuation effectiveness and improves downstream task performance, while significantly reducing computational cost, achieving up to 14× fewer iterations to convergence and 4× speedup in end-to-end execution. \end{abstract}

\begin{abstract} Designing intra-operator parallel patterns for large-scale Transformer training remains labor-intensive because it requires jointly coordinating tensor partitioning, local computation, and communication under hardware constraints. Existing automatic parallelization systems commonly select from predefined pattern libraries, limiting their ability to explore fine-grained execution alternatives. We present OPERA, a reinforcement learning framework for automatically generating executable intra-operator parallel patterns.OPERA formulates parallel pattern generation as a Markov Decision Process in which the environment tracks tensor placements, computation progress, and feasible computation and communication actions. A Transformer decoder encodes the history of layout and execution decisions, while its policy and value heads guide Monte Carlo Tree Search toward promising feasible trajectories. The resulting trajectories are reused to train the policy-value network, enabling the agent to progressively refine its decisions. A distributed program translator converts completed trajectories into executable programs and evaluates their numerical correctness, runtime, and resource consumption. Experiments on an eight-GPU NVIDIA A100 platform using BERT-Large and workloads derived from LLaMA and GPT configurations show that OPERA recovers 9 established intra-operator parallel patterns and generates additional feasible execution trajectories from primitive actions rather than selecting fixed templates. \end{abstract}

Training large Transformer language models is dominated by the cost of synchronous parameter updates: every block of weights is moved at every step using a gradient computed before any block has moved. The resulting trajectory is wasteful, because per-block updates are evaluated on a frozen snapshot of the rest of the network that the next step will already invalidate. Second-order methods such as Shampoo and K-FAC address this spatially through preconditioning, and sharpness-aware methods such as SAM trade compute for sharper minima, but both leave the temporal coupling between blocks untouched. We introduce \textbf{Gradient Interaction Graph Optimization (GIGO)}, a meta-optimizer that estimates a sparse directed graph of cross-block gradient dependencies online, partitions each step into contiguous sequential stages along this graph, and refreshes downstream gradients using a stale-activation approximation. The framework wraps any base optimizer, adds one cached boundary activation per stage ($\sim$32\,MB at $T_{\mathrm{seq}}{=}1024$, scaling linearly with sequence length), and is compatible with mixed precision and data parallelism. On GPT-2 Small through XL (124M--1.5B), GIGO-Adam reaches the same OpenWebText/C4 perplexity in 32.0--34.0\,\% fewer optimizer steps than synchronous AdamW --- a lead of 8--22 percentage points over the strongest 2024-vintage baselines (SOAP, Muon, Sophia, Lion, GaLore) on the Small comparison cohort --- with a per-step overhead of 32--38\,\% and a net wall-clock benefit ranging from $+$5.2\,\% (Small) to $+$16.2\,\% (XL) across the four sizes tested. The benefit vanishes or reverses where the method's preconditions do not hold: a synthetic block-diagonal control (SRR\,$=$\,0.0\,\%), the overhead-dominated Penn Treebank stress test ($-$18.5\,\% wall-clock), and a weak-coupling small-vision-Transformer diagnostic ($-$12.4\,\% wall-clock on CIFAR-100). Ablations decompose the mechanism: the partial-re-forward refresh primitive alone (random partition) accounts for 9.4\,\% SRR, the interaction-graph schedule contributes a further $\sim$22\,pp, and the layer-contiguity constraint controls overhead. Our convergence theorem matches the standard Adam $O(T^{-1/4})$ rate uniformly in graph-estimation quality, providing a safety guarantee; the empirical step-reduction is reported as an empirical regularity not formally predicted by the theory.

Quantum Neural Networks (QNNs) possess high expressivity via rich unitary transformations, offering abundant representational capacity while maintaining a parameter-effective architecture. In classical deep learning, the high expressivity of deep neural networks (DNNs) motivates the Strong Lottery Ticket Hypothesis (SLTH), which suggests that a randomly initialized DNN contains sparse subnetworks that achieve high performance immediately or serve as a warm start. On the quantum side, a key question is: \textit{Does a strong lottery ticket exist in QNNs?} Our answer is, \textit{Quantum-Strong Lottery Ticket Hypothesis (Q-SLTH)}, i.e., the first theoretical analysis for the existence of high-performing subnetworks in random scratch of QNNs. We validate our theory by proposing a quantum edge-popup algorithm to find strong subnetworks and empirically identifying winning tickets in randomly initialized QNNs on the MNIST and Fashion-MNIST benchmarks. For practice, Q-SLTH suggests a way to compress QNNs, allowing them to use fewer gates and thereby mitigating quantum-noise-induced degradation that is amplified across multiple layers of quantum gates.

Deep CNNs have achieved significant performance across a variety of computer vision tasks. However, their effectiveness comes at a cost of overparameterization. This leads to extensive computation and memory requirements, thereby limiting practical deployment in resource-constrained environments. A growing body of evidence suggests that many of these parameters are redundant, motivating the need for compression strategies that go beyond parameter-level heuristics. To overcome these challenges, we propose a layerwise, similarity-based filter-sharing framework for optimizing parameters of deep models by identifying and exploiting redundancy among filters. By leveraging representational similarity of the filters measured via Centered Kernel Alignment (CKA), our approach enables efficient filter-sharing over layers, significantly reducing the number of parameters, FLOPs and overall memory footprint. We form clusters of filters by measuring their feature-map similarity scores on the dataset, enabling compression decisions guided by representational behaviour rather than raw parameter values. We further use an entropy- and energy-based scaling mechanism that adaptively determines the number of shared filters per layer. Additionally, we extend the framework with a class-conditional formulation, enabling the construction of specialized compressed sub-models tailored to different subsets of tasks. This provides insight into how representational redundancy varies with task difficulty and allows the model to allocate capacity in a data-dependent manner. We conduct extensive experiments over VGG16, AlexNet, ResNet and MobileNetV2 models across standard benchmark datasets. The results demonstrate the effectiveness of our proposed method in achieving substantial parameter and FLOP reductions, while maintaining competitive accuracy. We compare our framework against several state-of-the-art compression methods from the literature. Our proposed framework offers a simple and effective approach to exploiting redundancy in deep networks, with practical implications for efficient deployment and adaptive model design.

Two-bit Post-Training Quantization (PTQ) offers an attractive path to deploying Large Language Models under strict memory budgets, yet scalar quantization suffers from catastrophic performance collapse at this extreme bit-width. While Vector Quantization (VQ) methods resolve the accuracy issue, existing approaches suffer from practical inefficiencies: complex optimization pipelines result in hours-long quantization, and some methods exhibit severely limited inference throughput. We propose EPTQ, a 2-bit PTQ framework available in two complementary variants (E8 and FE8), achieving high accuracy and practical efficiency through three contributions: (1) \textbf{Factored $E_8$ (FE8) lattice quantization}, which decomposes the $E_8$ codebook into a compact 4\,KB structure that fits in the GPU L1 cache, enabling fast codebook lookup at inference time; (2) \textbf{Lightweight scale normalization} via row/column scaling vectors ($\mathbf{g}$, $\mathbf{h}$), which aligns weight distributions to maximize $E_8$ quantization efficiency, requiring only element-wise per-token scaling at inference; and (3) \textbf{Adaptive critical weight preservation} via knee-point threshold detection, which more accurately identifies Hessian-sensitive weights than approaches relying on fixed global criteria, consistently improving perplexity across models, with zero bit-overhead by encoding the preservation mask into the existing codebook index space. Evaluated across multiple LLMs including Llama-2 and Llama-3 families, EPTQ (E8) achieves accuracy competitive with or surpassing QTIP on Llama-2 models, while EPTQ (FE8) surpasses competing VQ methods in decode throughput on most models, achieving up to 38\% faster inference than FP16, with both variants reducing quantization time by over 80\% compared to competing approaches.

As language models continue to grow in size due to their success in a wide range of applications such as conversational agents and coding, there is a pressing need to make them more accessible. A popular technique is to reduce their memory footprint by post-training quantization, where we use fewer bits (e.g., 4--8 bits) to represent the model without retraining it. However, quantizing language models remains a challenging task, especially in the ultra-low-bit setup (e.g., 2 bits). In this paper, we propose InvarExplore, a unified framework that systematically explores different kinds of model invariance to maximize quantization performance. Importantly, InvarExplore features a discrete search algorithm that enables us to explore permutation invariance, which is under-studied as it cannot be optimized with gradient-based methods. Such exploration allows us to take advantage of the synergy between each type of invariance. Our results and in-depth analyses show that InvarExplore is compatible with existing state-of-the-art integer quantization methods, achieving an add-on performance improvement over strong competing methods.

Analyzing high-velocity data streams is critical for tasks like real-time monitoring, anomaly detection, and performance analysis, driving interest in sketches, probabilistic data structures that efficiently summarize fast streams. Traditional sketches often use large, fixed-size counters, leading to inefficient memory usage in skewed data distributions. Recent solutions, such as hierarchical structures and dynamic counters, attempt to improve memory efficiency but struggle to balance this with detection accuracy. In this work, we present Nino-Sketch, a memory-efficient and versatile sketch for data stream analysis. Nino-Sketch employs a flexible filter to exclude cold items (e.g., those with low frequency or persistence) while focusing on tracking hot items using variable-size counters. When a counter overflows, it reallocates memory by sacrificing the slot with the smallest counter value, enhancing the tracking of hot items. By using variable-length fingerprints (a short identifier derived from an item's key using hashing techniques) for different item types, Nino-Sketch further optimizes memory usage. Experimental results on various datasets show that Nino-Sketch surpasses existing methods, reducing relative error by up to 11.17× over prior baselines under tight memory conditions (e.g., 16KB) while retaining competitive update throughput. We further deploy Nino-Sketch as a synthesizable RTL design on a Xilinx Kintex-7 FPGA, confirming that the bit-level layout is implementable on reconfigurable hardware.

We consider the problem of building space-efficient indexes for static key-value datasets with highly skewed value distributions. A standard technique for this task is to augment a compressed static function (CSF) with an approximate set membership prefilter to handle a dominating value and break the one-bit-per-key barrier. However, we observe that prior approaches decide when to apply such a prefilter using idealized cost models or data-driven heuristics, and, as a result, can recommend using a prefilter in regimes where it actively increases space usage. We introduce AutoCSF, a principled algorithm for combining CSFs with prefiltering that comes with a provable safety guarantee: with high probability, it does not recommend a filter when doing so would harm space usage. Our key insight is that while modeling the absolute cost of each index appears intractable, the difference in cost between a CSF and a filter-augmented CSF admits clean analytical bounds via the optimality of Huffman codes. We derive a lower bound on this difference whose sign yields a provably safe decision criterion, along with a complementary upper bound on the maximum achievable savings from prefiltering. Because AutoCSF depends only on the dominant value fraction and a filter cost function, it applies to any approximate set membership structure with a known cost-versus-false-positive-rate tradeoff. We also validate our bounds empirically across three synthetic value distributions and four real-world genomics workloads, and show that the decision criterion never recommends a space-increasing filter while its parameter selection costs at most 0.02 bits per key more than an exhaustive search.

Graph diffusion models have emerged as a promising approach for generating complex networks, yet they typically face two major drawbacks. First, terminating the forward process in pure Gaussian noise eliminates the intrinsic structural signatures of the original network, which leads to suboptimal generation quality. Second, the absence of constraints along the diffusion trajectory causes progressive degradation of topological features, ultimately resulting in complete structural collapse. To overcome these challenges, we introduce Null-Model-Constrained Graph Diffusion (NMC-GD), a principled framework incorporating specialized designs for graph generation. Specifically, we substitute the conventional isotropic prior with a null-model distribution as the new forward diffusion endpoint, explicitly preserving critical statistical properties of the network. In addition, we develop a null-model-guided continuous-time stochastic differential equation (SDE) and design a directional noise term that enforces a global structural constraint on the diffusion trajectory, striking a balance between stochasticity and structural invariance. Extensive experiments on three public datasets demonstrate that NMC-GD achieves superior generation quality and computational efficiency.

Modeling 3D dynamical systems is a fundamental challenge in the physical and engineering sciences, where Equivariant Graph Neural Networks (EGNNs) have emerged as a powerful paradigm by incorporating geometric symmetries. However, these models are fundamentally constrained by their discrete-time, Markovian framework, which neglects long-range temporal correlations and inevitably leads to error accumulation in long-horizon forecasting. To address this limitation, we introduce the Equivariant Graph Neural Ordinary Differential Equation (EG-NODE), a novel framework that directly learns the continuous-time evolution laws of physical systems. Instead of predicting discrete future states, EG-NODE leverages an equivariant GNN as its core to directly model the ordinary differential equation governing the system's instantaneous rate of change the physical laws of motion thereby natively preserving SE(3) symmetry within the learning process. This continuous-time paradigm enables high-precision predictions at arbitrary time points and allows for the use of adaptive step-size solvers to dynamically balance computational efficiency and accuracy. Extensive experiments on N-body, molecular, and fluid dynamics benchmarks demonstrate that EG-NODE significantly outperforms existing discrete models in long-horizon prediction accuracy and effectively suppresses error propagation. Our work establishes a more fundamental, first-principles-based paradigm for learning continuous physical laws from data. Our code is available at https://anonymous.4open.science/r/egno_egnn-436D/.

Graph Transformers show immense potential in graph representation learning, yet adaptively capturing complex topological dependencies remains a critical challenge. This paper proposes AdaGPS, a novel adaptive architecture featuring three core improvements. First, the introduced GatEdgeformer module injects edge feature information through a dynamic gating mechanism, enhancing the global attention's perception of complex graph structures. Second, dynamically fusing local structure perception with global attention interaction effectively filters structural noise and accurately extracts long-range dependencies. Third, dynamically fusing historical representations from all previous layers enables the deep AdaGPS architecture to fully utilize rich positional and structural encodings preserved across preceding layers. Extensive experiments strongly demonstrate that the injection of edge features and historical information effectively improves model accuracy. Furthermore, the gated dynamic fusion of local and global features significantly enhances cross-domain generalization capability. Systematic ablations confirm robustness and stability, providing a highly expressive new paradigm for graph representation learning and potential computing power network applications.

Edge classification, a crucial task for graph applications, remains relatively under-explored compared to link prediction. Current methods often overlook the potential causal influences of node features on edge features, leading to a loss of relevant prior information. In this work, we present an empirical exploration using the Causal Edge Classification Framework (CECF). Unlike conventional causal inference methods, CECF is the first framework to apply causal inference principles to the edge classification task and to explore modeling edge features as a high-dimensional treatment within a causal framework. Based on the node embedding of Graph Neural Network (GNN), CECF seeks to learn a balanced representation of high-dimensional edge features by mitigating the potential influence of node features. Then, a cross-attention network captures the complex dependencies between node and edge features for final edge classification. Extensive experiments demonstrate that CECF not only achieves superior performance but also serves as a flexible, plug-and-play enhancement for existing methods. We also provide empirical analyses, offering insights into when and how this high-dimensional causal modeling framework works for the edge classification.

Graph neural networks (GNNs) have achieved remarkable success in node classification tasks, but their performance often varies significantly across classes. Some classes benefit from highly accurate predictions, while others perform relatively poorly. Recent advances in imbalanced node classification have provided partial remedies, but often lead to a seesaw effect: improvements in the minority class come at the expense of the majority class. This phenomenon reflects a deeper problem of inter-class interference, where optimizing for some classes during training interferes with or inhibits learning in other classes. This imbalance is closely related to representation bias, where limited representational capacity is unevenly distributed across classes. To address this issue, we propose a graph balancing learning framework based on the structural rebalancing theory. Extensive node classification experiments demonstrate that our framework is applicable to various existing GNN models and effectively addresses the seesaw effect of GNNs. The code is available at https://github.com/linlin-oss/SSER-main.

Graph Neural Networks (GNNs) have shown promising performance in learning from graph-structured data, but at the cost of prohibitive computational overhead due to a series of graph-scale operators. To reduce redundant graph operations, previous studies attempt to sparsify the graph by pruning or sampling graph edges, which enables popular GNN deployments in realistic tasks. However, these methods are mostly empirical, lacking a theoretical framework for bounding approximation precision and guiding sparsification policy. In this work, we present the first approximation theory for GNN sparsification that unifies both pruning and sampling paradigms, directly characterizing the representation error deviation in a layer-wise manner. Our theoretical bounds reveal that the pattern of error accumulation throughout iterative graph operators varies with key GNN model properties. Based on our framework, we further derive an optimal sparsification policy that minimizes the layer-wise approximation error under fixed computational constraints. Experimental evaluations confirm our theoretical findings on layer approximation bounds. Remarkably, our proposed optimal sparsification allocation reduces the representation error by up to $10\times$ and retains better prediction accuracy compared to canonical sampling strategies. The code is available at https://github.com/Rechard16/QuOS/.

Distributed graph neural network (GNN) training is essential for large-scale graph learning, but existing mini-batch frameworks still suffer from substantial inter-partition communication overhead and indirect, coarse-grained GPU memory control. In distributed mini-batch training, remote sampled neighbors cause frequent feature fetching, while fixed seed-node batch sizes cannot accurately reflect the computation graph expansion induced by multi-layer sampling. This paper presents BPB-GNN, a resource-efficient distributed GNN training system that jointly optimizes inter-partition feature access, hierarchical storage management, and mini-batch construction. First, BPB-GNN introduces a training-centric boundary-aware prefetching strategy to localize high-value remote neighbor features associated with boundary nodes in training-node receptive fields. Second, it uses a DRAM/NVMe SSD hierarchy with priority-aware placement and dynamic migration to manage prefetched features under limited DRAM capacity. Finally, it constructs seed-node batches according to estimated sampling-induced load costs rather than equal seed-node counts, thereby balancing batch-level workloads and reducing peak GPU memory usage and its inter-batch fluctuation. These resource-efficient mechanisms may also improve the efficiency of large-scale GNN computation in computing power networks and JointCloud environments. Experiments on real-world graph datasets show that BPB-GNN achieves up to 5× speedup over the baselines while reducing peak GPU memory usage by up to 36%.

Classical AC power flow assumes that nodal active and reactive power injections are fully specified. In practice, however, reactive power at some PQ buses may be unavailable due to limited observability or implicit local control. In this paper, we study power flow calculation with partially observed reactive power, where network topology, bus types, active power, and only a subset of reactive-power injections are available as inputs. Specifically, we propose \textsc{Q-Flow}, a graph-based framework for predicting voltage states from such incomplete inputs. Q-Flow combines a Hierarchical Information Fusion (HI-Fuse) with a Multi-Order Gated Kolmogorov–Arnold Network (MoG-KAN), allowing the model to exploit local grid structure, bus-type information, and system-level context while modeling the nonlinear relationship between power injections and voltage states. AC power-flow equations are further incorporated as a physical consistency constraint during training. We evaluate Q-Flow on IEEE 39-, 118-, and 300-bus systems under fixed and sample-dependent reactive-power masks, as well as the complete-Q setting. Q-Flow consistently outperforms representative graph-learning baselines and maintains low voltage-state errors as reactive-power observability decreases.

Recommender systems are often evaluated using observable behavioral signals as surrogate relevance measures of user experience. However, in well-being and mental health support applications, user experience is inherently non-observable. To address this limitation, we introduce a multimodal physiology-driven framework for evaluating recommendation rankings in art therapy using functional near-infrared spectroscopy (fNIRS) and eye-gaze signals. We define a \emph{physiological utility} scalar that captures item-level internal response by combining prefrontal hemodynamic responses with attention and engagement dynamics derived from gaze behavior, enabling direct comparison between system-generated rankings and rankings induced by internal user responses. We show that fNIRS-derived utility exhibits moderate alignment with system rankings (Spearman’s $\rho \approx 0.40$, pairwise accuracy $\approx 0.69$), with inter-hemispheric functional connectivity showing the strongest alignment among the evaluated neural descriptors. Gaze-derived utility yields stronger observed alignment ($\rho \approx 0.55$, pairwise accuracy $\approx 0.75$), particularly for attention continuity and engagement-related representations. Multimodal integration yielded the highest observed agreement between recommendation rankings and physiology-derived utility ($\rho \approx 0.61$, pairwise accuracy $\approx 0.77$). Our findings provide exploratory evidence that multimodal physiological signals may offer complementary information for evaluating recommendation rankings in outcome-sensitive art therapy.

Large language models (LLMs) are increasingly used in recommender systems for user modelling, ranking, and simulation. Among these applications, LLM-based user simulators provide a flexible way to evaluate recommender systems without relying on live user feedback. However, existing simulators typically reduce user behaviour to static preference matching based on fixed user profiles, overlooking the fact that real decisions may also reflect recent preference shifts and exploratory behaviour. To address this limitation, we propose \textbf{IntRecSim} (\textbf{Int}ent-aware \textbf{Rec}ommendation \textbf{Sim}ulation), a plug-in behavioural control framework for LLM-based user simulation. IntRecSim introduces a lightweight intent mechanism that guides whether the simulated decision should follow long-term preferences, recent behavioural evidence, or exploratory behaviour. Since standard recommendation datasets lack explicit intent annotations, we construct pseudo-intent supervision from historical interactions using hybrid category and semantic preference signals. The inferred intent is injected into the LLM simulator as a dynamic control signal, making the framework compatible with different profile-generation backbones. Experiments on three benchmark datasets, covering discrimination and ranking tasks show that IntRecSim improves simulation performance in most evaluated settings. Further analysis highlights the importance of explicit behavioural control for modelling realistic user decisions in LLM-based recommender simulation.

Reliable evaluation of recommender systems (RSs) often requires user studies or online interaction logs, both of which are costly to obtain for new or low-resource platforms. User behaviour simulation offers a scalable alternative, but existing simulators are typically trained and deployed within a single domain. Thus, they cannot easily reuse the same user agent across domains and often conflate stable user traits with domain-specific interaction evidence. In this paper, we study cross-domain user behaviour simulation: given a user agent learned from a source domain, simulate the same user's behaviour in a target domain. We propose TMemAgent, a transferable memory-based user simulator. The key idea is to treat memory as an actively managed behavioural representation rather than a passive store of past interactions. TMemAgent maintains two complementary memory banks: a transferable memory bank that captures stable user characteristics across domains, and a domain-specific memory bank that preserves local contextual evidence. A memory learning and management (MLM) policy, trained with an actor-critic objective, dynamically decides when to retrieve, add, reflect, update, or forget memory entries. Experiments on three Amazon domains show that TMemAgent improves user-agent prediction performance, yielding gains of around 12% in same-domain settings and around 14% in cross-domain settings compared with the strongest baseline. Moreover, the transferred agents preserve behavioural diversity close to the ground truth, demonstrating that explicit MLM supports scalable, transferable, and behaviourally realistic user simulation for RS evaluation.

User behavior simulation is widely used to evaluate interactive information retrieval systems, but classical state-based approaches (e.g., Markov models) have limited ability to incorporate contextual information relevant for decision-making. We address this limitation by introducing a feature-conditioned Markov-style user model, in which transition probabilities are modeled as functions of positional, content-based, and interaction-derived features, enabling context-aware decision making while preserving the structural simplicity and computational efficiency of state-based models. Applying a multi-level framework that assesses predictive fit and behavioral fidelity, we analyze how different sources of contextual information contribute to realistic user simulation across multiple datasets, search settings, and feature configurations. Our results show that incorporating contextual features improves the models' ability to reproduce key aspects of real user interactions, but that their effectiveness hinges on search scenario and modeling objective. Instead of a \enquote{one-size-fits-all} solution, effective simulation requires task- and setting-specific feature selection. Our framework provides a practical and interpretable basis for making these choices.

\begin{abstract} Large language models (LLMs) are increasingly used for information seeking, where users find, compare, and evaluate information through dialogue. In this role, the assistant does more than retrieve or generate content: it shapes how users articulate constraints, ask follow-up questions, verify claims, and decide when an answer is sufficient for action. Yet little is known about how user personality, assistant personality, and task context jointly influence these interactions. We examine personality as a controllable variable in conversational information seeking and study its effects on user behaviour and interaction quality. We conducted a controlled within-subject study in which assistant personality and task type were experimentally varied, while participant personality was measured using Big Five scores. Twenty-six participants each completed three information-seeking tasks under three assistant personality conditions: extraverted, conscientious, and neutral. Tasks covered exploratory travel planning, comparative smartphone shopping, and verification-sensitive health and diet information seeking. Data included conversation logs, behavioural traces, post-interaction questionnaires, an exit questionnaire, and Big Five measures. The assistant conditions were behaviourally distinct: the extraverted assistant produced longer turns, the conscientious assistant elicited higher user word share and more turns, and the neutral baseline fell between them. The strongest effect was a task-by-assistant interaction on trust and delegation, with preferred styles varying by task. No global winner emerged, but participants strongly preferred style choice or adaptation. These findings position assistant personality as a context-sensitive interactional design variable rather than a globally optimisable system property. \end{abstract}

Food recommender systems are increasingly expected to encourage healthier, more sustainable choices, often through sustainability-aware explanations that blend transparency with persuasive messaging. Yet little is known about how users' cultural backgrounds and cognitive styles shape their responses to these explanations. In this paper, we investigate the influence of cultural background and cognitive style on users' behavioral decisions and subjective evaluations of sustainability-aware explanations in food recommender systems. We conducted a user study using a web-based food recommendation prototype incorporating three Cialdini-inspired persuasive explanation strategies: Authority, Commitment, and Social Proof. The study compared Asian and European participants and further examined differences across cognitive-style groups. The results show that cultural background significantly influenced both recipe-selection behavior and users' perceptions of explanation quality, whereas cognitive style showed comparatively smaller effects that mainly affected subjective explanation evaluations. On the other hand, the persuasive explanation strategies themselves produced relatively stable behavioral effects across cultural and cognitive-style groups, with no significant interaction effects after multiple-comparison correction. These findings provide practical implications for designing food recommender systems that effectively balance persuasive support, transparency, and cultural adaptability to encourage healthier and more environmentally responsible consumption behaviors.

Conversational recommender systems (CRSs) infer user preferences from dialogue contexts, but they suffer from severe data sparsity in both dialogue and entity spaces. Dialogue data are linguistically diverse and open-ended, making it difficult to generalize across varied expressions. Meanwhile, existing CRS models often rely on large knowledge graphs, where only a small fraction of entities receive effective supervision during training, leaving the majority under-trained or entirely unseen at inference time. To address these challenges, we propose DACRS, a novel CRS framework consisting of three modules: Dialogue Augmentation, Knowledge-Guided Entity Modeling, and Dialogue–Entity Matching. The Dialogue Augmentation module adopts a two-stage augmentation pipeline to enrich dialogue contexts and improve robustness to linguistic variation. The Knowledge-Guided Entity Modeling module leverages knowledge graphs through entity substitution and an entity similarity constraint to enhance representation learning for sparsely supervised and unseen entities. Finally, the Dialogue–Entity Matching module integrates dialogue representations with mentioned entity embeddings via dialogue-guided attention aggregation, yielding user representations that capture both explicit and implicit preferences. Extensive experiments on two public benchmark datasets demonstrate that DACRS consistently outperforms state-of-the-art conversational recommender systems.

A safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53% of safety rules after one compaction round and 10\% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its own retention policy. Three deterministic operators implement this triage across the three context-management operations: TypeCompact rewrites items in place under per-type fidelity, TypeDecompose partitions a topic too large to compact safely, replicating in-scope safety rules across partitions, and TypeRetrieve fetches items from external storage with in-scope rules pinned ahead of relevance. On five public corpora, TypeCompact preserves 2-4$\times$ more safety rules than the strongest single-shot LLM compactor at every ratio, with 96% recall over five rounds. TypeDecompose reaches 0% locality violations against 93% under uniform partitioning. TypeRetrieve reaches 100% recall@50 against 73% for the best single-shot LLM retriever. On three downstream behavioral benchmarks, we outperform the production Sonnet compactor on medical compliance (paired McNemar $p < 10^{-8}$ on preservation, $N = 200$), the full-policy and hierarchical baselines on retail task pass rate ($p < 0.01$, $N = 115$), and the hierarchical compaction on the airline domain ($p = 0.024$). We release AgentArtifactCorpus (396{,}934 agent configurations from 54{,}628 public GitHub repositories), the classifier, and the reference implementation.

For long-horizon LLM agents, forgetting is not a housekeeping detail: it determines whether the agent preserves commitments, updates stale beliefs, and reasons coherently under finite memory. Yet existing policies mostly score each memory in isolation, so they cannot capture a central fact of agent interaction: a later event can revive or disambiguate an earlier memory. This mismatch makes budgeted forgetting a structural bottleneck for reliable agent memory. We propose \textsc{H-Mem}, a deceptively simple retention controller that turns forgetting into an event-coupled process. \textsc{H-Mem} first partitions memories into semantic clusters, allocates budget across clusters to avoid topic monopoly, and then runs cluster-local Hawkes excitation so that new related memories reinforce useful predecessors with diminishing returns. Bounded base intensity protects isolated but potentially important memories before relational evidence accumulates. This design confines propagation to semantically valid neighborhoods while keeping retention decisions auditable, causal, and budget-aware. We formalize an indistinguishability barrier for item-wise forgetting: when two memories share the same self history but differ in neighborhood evidence, any isolated scorer must make the same decision and can incur unavoidable regret. We also prove feasibility, bounded intensity, noise-stable influence, anti-monopoly quota behavior, optimal partition-matroid projection, monotone quota sensitivity, and end-to-end complexity. Experiments on LoCoMo and LongMemEval-S show that \textsc{H-Mem} consistently improves QA and evidence preservation; on LoCoMo at 25\% memory, it raises Qwen QA from $42.6$ to $47.4$ and Evidence Survival@B from $41.3$ to $61.1$ over the strongest baseline under a matched memory pipeline.

Reinforcement learning with verifiable rewards (RLVR) is a practical way to improve the reasoning and decision-making abilities of large language models, but its training signal is often sparse. When a policy faces prompts beyond its current ability, every sampled trajectory in a group may fail the verifier, resulting in an all-zero group. In this case, group-relative policy optimization gives little useful gradient even when the prompt contains valuable learning content. This capacity-difficulty mismatch is especially costly for smaller models used in knowledge-intensive and web-interaction tasks. We present \name, a difficulty-aware training framework that augments on-policy RL with conditional guidance from a larger reasoner. \name uses the standard GRPO branch when at least one sampled response succeeds, and switches to a supervised objective over verified teacher trajectories only when the whole student group receives zero reward. The method can turn otherwise wasted all-zero groups into targeted learning opportunities while keeping policy exploration on prompts the student can handle. We formalize the switching criterion, teacher selection objective, mixed RL/supervised loss, and compute overhead. We evaluate \name on WebShop, GAIA, and WebWalkerQA using Qwen2.5-7B-Instruct, Qwen3-4B, and Qwen3-8B policies. The results show consistent gains over matched GRPO baselines, with large improvements on hard web traversal questions for the two smaller policies. Ablations and diagnostics further show that the gains come from the proposed all-zero recovery mechanism.

Autonomous agents powered by Large Language Models (LLMs) can automate complex web tasks, but their reliability is limited by uncertain action outcomes: the consequence of interacting with a web element is often known only after execution. Existing sequential methods are vulnerable to error propagation after wrong decisions, while tree-search approaches rely on speculative pre-execution scoring and incur substantial exploration cost. We introduce WebTactix, an evidence-based parallel probing framework that shifts web navigation from speculative prediction to grounded decision making. Rather than scoring actions before execution, WebTactix executes multiple candidate plans in parallel, observes their successor pages, and compares the resulting evidence to select the next branch while accounting for future continuation. This design enables the agent to explore multiple alternatives without committing prematurely to uncertain decisions or incurring the overhead of extensive tree search. Experiments on WebArena and Online-Mind2Web demonstrate consistent improvements in both task success and efficiency, with WebTactix achieving 76.8% success on WebArena while reducing average completion time by 56.2% relative to the strongest tree-search baseline.

Agentic Retrieval-Augmented Generation (RAG) has become a promising paradigm for multi-hop question answering, where a reasoning model iteratively issues queries to a retriever and incorporates newly retrieved context into subsequent reasoning steps. While this iterative process can improve final answer quality, current evaluations of agentic RAG largely focus on end-to-end outcomes and provide limited visibility into how a model's answer state changes during generation. In this work, we introduce an in-trajectory probing framework to study intermediate answer states in agentic RAG. Specifically, after each retrieval-reasoning iteration, we force an agentic model to stop reasoning and generate an intermediate answer based on its current state. This allows us to define two iteration-level measures: partial answer quality at each iteration, and partial utility as the change in partial answer quality across iterations. Our analysis across multi-hop QA benchmarks reveals that partial answer quality often plateaus before natural termination, with many later iterations contributing only small measurable improvements. Accordingly, we formulate two prediction tasks, partial answer quality prediction and partial utility prediction, and study trajectory-derived signals from intra-iteration, inter-iteration, and query-iteration perspectives. Experiments show that partial answer quality is more predictable than partial utility, with supervised models achieving Pearson's $r$ above 0.43 for quality prediction. Finally, using predicted answer quality and utility for early stopping reduces average iteration count by about 11\% while preserving about 98\% of the final answer quality achieved by natural stopping.

Recent search agents for multi-hop reasoning often fail by either retrieving incomplete evidence or reasoning over irrelevant portions of the retrieved content, leading to a retrieval-reasoning boundary shift. We propose R$^2$-Searcher, a novel framework that explicitly explores and calibrates the retrieval and reasoning boundaries via fine-grained, query-token-guided evidence modeling and post-retrieval reflection. Specifically, R$^2$-Searcher: (1) constructs fine-grained reasoning contexts by extracting precise facts from retrieved content based on query token semantics (e.g., subjects, actions, temporal markers, and degree modifiers), thereby guiding the attention of search agent; (2) introduces a retrieval reflection mechanism that evaluates and corrects boundary deviations after each retrieval step, guiding the generation of improved queries grounded in the extracted reasoning contexts; and (3) employs an end-to-end reasoning-reflection-guided reinforcement learning algorithm, R$^2$PO, which jointly optimizes both boundaries through a tree-based exploration of reasoning regions and reflections. Our method significantly enhances the quality of both retrieval and reasoning, establishing an iterative loop where retrieval and reasoning mutually enhance each other. Extensive experiments on seven complex multi-hop QA benchmarks demonstrate that R$^2$-Searcher significantly outperforms state-of-the-art agentic search methods in answer accuracy and retrieval-reasoning quality. Ablation studies further confirm the critical role of retrieval-reasoning boundary calibration. Implementation code is available for reproducibility: https://github.com/Applied-Machine-Learning-Lab/CIKM2026\_R2-Searcher.

Synthesizing informative commercial reports from massive and noisy web sources is critical for high-stakes business decisions. Although recent deep research agents (DRAs) achieve notable progress, their reports remain limited in quality, reliability, and coverage. These mainly stem from ambiguous intents that cause search drift, retrieved web content that rapidly saturates the context window, and single-pass synthesis that limits report comprehensiveness. In this work, we propose \Method, a cognitive deep research agent that emulates commercial analysts to synthesize expert-level reports. \Method\ first probes fine-grained commercial intent to establish a structured outline, then recursively explores web sources and distills validated evidence into research memory to preserve context efficiency. Meanwhile, the research memory and outline continuously co-evolve, refining the report structure to avoid rigid initial planning. Finally, \Method\ iteratively synthesizes the report based on the evolving outline and accumulated evidence. Together, these designs enable reliable and context-efficient long-horizon commercial deep research. To rigorously evaluate commercial DRAs, we further construct \Evaluation, comprising 200 real-world commercial tasks and a holistic evaluation framework covering report quality, reliability, and coverage. Extensive experiments demonstrate that \Method\ consistently outperforms leading proprietary and open-source DRAs, while ablations verify the effectiveness of each module and further analyze the challenges they address. We expect this work to advance the development of commercial deep research agents. The implementation of \Method\ and the \Evaluation\ data are publicly available \footnote{https://github.com/ustc-ai4science/Mind2Report}.