Mixture-of-Experts (MoE) models are central to scaling Large Language Models (LLMs), but stateless and compute-intensive routing repeatedly re-explores expert assignments for similar inputs, causing computational redundancy and unstable behavior in web-scale applications like search and dialogue. We reframe expert routing as a retrieval-augmented process and propose the Retrieval-Memory Synergy Mixture-of-Experts (RMS-MoE), which integrates a Co-Activation Memory (CAM) to store and retrieve effective expert teams and a learnable, input-dependent gate to fuse retrieved priors with live routing predictions, enabling consistent expert coordination for semantically related inputs. Extensive experiments on web-scale QA and dialogue tasks show that RMS-MoE achieves a 26% latency reduction, a 2.7-point accuracy gain, and a 3.3% improvement in routing stability, demonstrating architectural memory as a principled path toward more efficient, stable, and scalable LLMs.
论文检索
输入标题、作者或关键词，从 954 篇学术成果中精准定位
Modern web interfaces increasingly support complex decision workflows, such as travel planning and multi-criteria selection, yet remain largely static and insensitive to users' moment-to-moment cognitive states during interaction. Travel planning, in particular, requires users to synthesize dispersed information under multiple constraints, making it a representative high-load interactive decision task. This study presents MACA (Multi-Agent Cognitive Adaptation), a framework that enables real-time cognitive adaptation in web-based decision environments by integrating hierarchical Monte Carlo Tree Search with a Planner–Critic–Executor multi-agent architecture. MACA continuously estimates users' emotional and attentional states using facial expression analysis (ResEmoteNet) and gaze stability tracking (MediaPipe), and uses these signals to regulate agent collaboration, reasoning depth, and feedback pacing during interaction. We evaluated MACA in a 2×2 within-subject study (N = 30) comparing Single versus Multi-agent and Fixed versus Adaptive configurations. Results show that the Multi-Adaptive condition significantly improved decision quality (F(3,116) = 2.96, p = 0.035) while reducing mental effort (F(3,116) = 2.82, p = 0.042), yielding a 10.7% gain in decision efficiency without increasing cognitive burden. These findings demonstrate that multimodal user-state sensing combined with cooperative multi-agent reasoning can enhance interactive web-based decision making while maintaining user well-being.
Self-explaining recommenders enhance user trust by providing justifications for their suggestions. Sequence-aware models have advanced the field by leveraging user interaction history to personalize recommendations and explanations. However, generative models often struggle with sparse data, producing repetitive or irrelevant explanations. This paper explores the optimal methods for infusing rich textual information from past user interactions directly into the item embeddings to feed a user reasoning path leading to personalized explanations. We conduct a comprehensive analysis of various techniques, including: (1) multiple text aggregation strategies to pool fine-grained attributed item opinions into user-aggregated item text representations; (2) several fusion mechanisms to combine text and collaborative modalities, from early fusion to a late fusion approach within the Transformer architecture; and (3) different training regimes for explanation generation. Experiments on three real-world datasets demonstrate which steps to follow in order to successfully leverage textual information into a sequence-aware explainable recommendation model and boost recommendation performance as well as explanation quality.
The proliferation of highly realistic deepfake videos threatens public trust and the integrity of digital information. However, detecting sophisticated deepfakes requires analysis beyond surface-level visual artifacts. We propose Harmonizing Action Units with Temporal-contextual Embeddings (HAUTE), integrating physiological muscle dynamics with holistic semantic context through adaptive attention mechanisms. HAUTE captures temporal Action Unit coordination patterns and high-level contextual embeddings, enabling the model to reveal synthesis-induced inconsistencies imperceptible to isolated modalities. Extensive experiments demonstrate state-of-the-art performance with strong cross-dataset adaptability, particularly on commercial tool-based high-quality deepfakes, advancing trustworthy content verification for web ecosystems.
Money laundering enables malicious actors to integrate illegal profits into the legitimate economy and has long been a central concern in financial regulation. Blockchain systems introduce new channels for laundering through decentralized, pseudonymous, and cross-border asset transfers. In this context, blockchain exploiters often rely on laundering to conceal fund origins and enable cash-out. While prior work has focused on detecting suspicious accounts or transactions, the behavioral patterns underlying laundering practices remain underexplored. This paper provides a behavioral perspective on post-exploit laundering on Ethereum. We use on-chain tracing to reconstruct token flows originating from exploiter-controlled addresses. We then define a set of behavioral metrics covering financial trajectories, temporal dynamics, structural topology, and value dispersion. Our empirical study reveals recurring patterns, including rapid fund movement, shallow transfer structures, and broad dispersion. These patterns exhibit measurable regularities, suggesting that behavioral dynamics could be leveraged to enhance existing laundering detection frameworks.
Accurate estimation of delivery time (EDT) is a critical factor in web e-commerce user experience. The pursuit of higher EDT accuracy has predominantly centered on designing increasingly complex model architectures. While valuable, this architecture-centric paradigm creates a tension between its high iteration costs and the industrial demand for agile deployment. This work, therefore, explores a complementary dimension: enhancing model performance by optimizing the learning process itself. We propose EDTF, a novel, plug-and-play composite learning framework that empowers existing models by augmenting their learning objective. EDTF first transforms the traditional regression problem into a structured ordinal classification task to address the training difficulties inherent in direct regression and preserve temporal order. It then introduces a cross-view consistency paradigm, decomposing the prediction task into two related views: the macroscopic end-to-end delivery time and the microscopic next-hop duration. By enforcing a self-supervised signal that aligns the sum of future next-hop durations with the overall EDT, our framework enables models to learn more robust temporal representations without extra features. Extensive experiments on a large-scale industrial dataset show that EDTF, as a plugin, consistently enhances performance and accelerates convergence across five diverse architectures. Critically, an EDTF-optimized model has been successfully deployed in a live production environment, demonstrating significant improvements over its predecessor. This work thus presents a validated and valuable new paradigm for the economical and efficient application of web services reliant on trajectory-based forecasting, from e-commerce to ride-hailing and food delivery.
General-purpose embedding models have demonstrated strong performance in text retrieval but remain suboptimal for table retrieval, where highly structured content leads to semantic compression and query–table mismatch. Recent LLM-based retrieval augmentation methods mitigate this issue by generating synthetic queries, yet they often rely on heuristic partial-table selection and seldom leverage these synthetic queries as supervision to improve the embedding model. We introduce CGPT, a training framework that enhances table retrieval through LLM-generated supervision. CGPT constructs semantically diverse partial tables by clustering table instances using K-means and sampling across clusters to broaden semantic coverage. An LLM then generates synthetic queries for these partial tables, which are used in hard-negative contrastive fine-tuning to refine the embedding model. Experiments across four public benchmarks (MimoTable, OTTQA, FetaQA, and E2E-WTQ) show that CGPT consistently outperforms the existing baselines with an average R@1 improvement of 16.54%. Under cross-domain evaluation, CGPT further demonstrates strong cross-domain generalization and remains effective even when using smaller LLMs for synthetic query generation. These results indicate that semantically guided partial-table construction, combined with contrastive training from LLM-generated supervision, provides an effective and scalable paradigm for large-scale table retrieval. Our code is available at https://github.com/yumeow0122/CGPT.
Behavioral patterns captured in embeddings learned from interaction data are pivotal across various stages of production recommender systems. However, in the initial retrieval stage, practitioners face an inherent tradeoff between embedding expressiveness and the scalability and latency of serving components, resulting in the need for representations that are both compact and expressive. To address this challenge, we propose a training strategy for learning high-dimensional sparse embedding layers in place of conventional dense ones, balancing efficiency, representational expressiveness, and interpretability. To demonstrate our approach, we modified the production-grade collaborative filtering autoencoder ELSA, achieving up to 10× reduction in embedding size with no loss of recommendation accuracy, and up to 100× reduction with only a 2.5% loss. Moreover, the active embedding dimensions reveal an interpretable inverted-index structure that segments items in a way directly aligned with the model's latent space, thereby enabling integration of segment-level recommendation functionality (e.g., 2D homepage layouts) within the candidate retrieval model itself. Source codes, additional results, as well as a live demo are available at https://github.com/zombak79/compressed\_elsa.
We provide the first anonymity-preserving algorithm for a centralized decision maker in linear bandit-based multi-user systems. Our algorithm employs successive elimination techniques for linear bandits to build an assignment multi-graph (from users to arms) along with a greedy matching algorithm that efficiently allocates the arms to users. We provide lower and upper bounds for this problem, showing that our algorithm is regret optimal up to a √(CK) factor.
Generative images have proliferated on Web platforms in social media and online copyright distribution scenarios, and semantic watermarking has increasingly been integrated into diffusion models to support reliable provenance tracking and forgery prevention for web content. Traditional noise-layer-based watermarking, however, remains vulnerable to inversion attacks that can recover embedded signals. To mitigate this, recent content-aware semantic watermarking schemes bind watermark signals to high-level image semantics, constraining local edits that would otherwise disrupt global coherence. Yet, large language models (LLMs) possess structured reasoning capabilities that enable targeted exploration of semantic spaces, allowing locally fine-grained but globally coherent semantic alterations that invalidate such bindings. To expose this overlooked vulnerability, we introduce a Coherence-Preserving Semantic Injection (CSI) attack that leverages LLM-guided semantic manipulation under embedding-space similarity constraints. This alignment enforces visual-semantic consistency while selectively perturbing watermark-relevant semantics, ultimately inducing detector misclassification. Extensive empirical results show that CSI consistently outperforms prevailing attack baselines against content-aware semantic watermarking, revealing a fundamental security weakness of current semantic watermark designs when confronted with LLM-driven semantic perturbations.
Attention mechanisms are essential to the success of Large Language Models (LLMs). In practice, models often overemphasize semantically low-value tokens, forming attention sinks while failing to capture truly informative tokens. Existing inference-time optimization methods mainly rely on static adjustments or attention redistribution, which often disrupt the correspondence between attention distribution and the actual semantics of the input, leading to a loss of semantic consistency and degraded performance. To address this problem, we propose PAOSC, a plug-and-play attention optimization model designed to maintain semantic consistency by dynamically adjusting attention. PAOSC employs a generator to identify informative tokens and a discriminator to optimize the generator via policy gradients based on confidence changes and loss fluctuations. Experiments on eight LLMs show up to a 9.68% improvement in the F1 score. On the constructed HTTP-RL dataset, PAOSC eliminates 18% of low-value tokens, improving inference efficiency while maintaining semantic consistency. Our code is available at https://github.com/ChangLi000/PAOSC.
To improve resource utilization in Cloud clusters and reduce operational costs, this paper proposes a two-level collaborative Vertical Pod Autoscaler (VPA) strategy for offline tasks in hybrid deployment environments. Offline tasks feature significant runtime variation and low periodicity—characteristics that make existing VPA strategies ineffective. The proposed approach combines coarse-grained adjustments (dynamically recommending resource settings based on cluster-wide utilization trends) with fine-grained adjustments (performing instance-level resource recalibration via sliding windows). It addresses the averaging perspective limitation of coarse-grained methods without requiring predictive models or prior knowledge, ensuring interpretability and operational simplicity. Production deployment verifies that the strategy elevates offline resource utilization to near-target levels, achieving an 8.62% improvement in offline resource utilization and a 41.5% increase in offline task deployment volume under Out-of-Memory (OOM) and computational constraints.
We consider the problem of graph sparsification while preserving the test accuracy of Graph Neural Networks (GNNs). Prior work in this area is often motivated by the Lottery Ticket Hypothesis, which aims to prune redundant edges. However, these sparsification approaches typically operate as black boxes and provide no justification for which edges are removed. In contrast, edge importance scores obtained from GNN explanation methods provide a principled and interpretable basis for sparsification. In particular, we show that Shapley value–based explainers such as GNNShap enable effective sparsification, allowing up to 80% of edges to be removed without degrading model accuracy. We show that Shapley values are well-suited for this task due to their robustness in identifying less influential edges, resulting in sparse yet faithful subgraphs that are efficient for downstream applications.
We propose a novel MLaaS Dataset Generator (MDG) framework that creates configurable and reproducible datasets for evaluating Machine Learning as a Service (MLaaS) selection and composition. MDG simulates realistic MLaaS behaviour by training and evaluating diverse model families across multiple real-world datasets and data distribution settings. It records detailed functional attributes, quality of service metrics, and composition-specific indicators, enabling systematic analysis of service performance and cross-service behaviour. Using MDG, we generate more than ten thousand MLaaS service instances and construct a large-scale benchmark dataset suitable for downstream evaluation. We also implement a built-in composition mechanism that models how services interact under varied Internet of Things conditions. Experiments demonstrate that datasets generated by MDG enhance selection accuracy and composition quality compared to existing baselines. MDG provides a practical and extensible foundation for advancing data-driven research on MLaaS selection and composition.
Inspired by advances in LLMs, reasoning-enhanced sequential recommendation performs multi-step deliberation before making final predictions, unlocking greater potential for capturing user preferences. However, current methods are constrained by static reasoning trajectories that are ill-suited for the diverse complexity of user behaviors. They suffer from two key limitations: (1) a static reasoning direction, which uses flat supervision signals misaligned with human-like hierarchical reasoning, and (2) a fixed reasoning depth, which inefficiently applies the same computational effort to all users, regardless of pattern complexity. These rigidity lead to suboptimal performance and significant computational waste. To overcome these challenges, we propose DTRec, a novel and effective framework that explores the Dynamic reasoning Trajectory for Sequential Recommendation along both direction and depth. To guide the direction, we develop Hierarchical Process Supervision (HPS), which provides coarse-to-fine supervisory signals to emulate the natural, progressive refinement of human cognitive processes. To optimize the depth, we introduce the Adaptive Reasoning Halting (ARH) mechanism that dynamically adjusts the number of reasoning steps by jointly monitoring three indicators. Extensive experiments on three real-world datasets demonstrate the superiority of our approach, achieving up to a 24.5% performance improvement over strong baselines while simultaneously reducing computational cost by up to 41.6%.
Multimodal disinformation may arise from the image, the text, or their cross-modal consistency, and these aspects can mislead either alone or in combination. However, existing research has not considered the full range of such combinations, leaving current methods vulnerable to unseen combinations. To address this gap, we propose a tri-axis formulation that explicitly separates image veracity, text veracity, and cross-modal consistency, yielding an 8-way space that fully enumerates all cases and turns multimodal disinformation detection, formerly a coarse, multi-label problem, into a well-defined, mutually exclusive classification task. To support this formulation, we construct OctantFake, an 8-way dataset sourced from 16 datasets with complete label coverage and broad class diversity. Experiments show that existing methods struggle on this challenging dataset. Building on this dataset, we introduce OctantAgent, an LVLM-based parallel framework with three modules, image check, text check, and consistency check, whose predictions are fused into an 8-way decision, achieving clear improvements over existing LVLM-based methods. Together, these contributions introduce an interpretable and exhaustive 8-way perspective that provides a solid foundation for analyzing multimodal disinformation.
Event forecasting is inherently influenced by multifaceted considerations, including international relations, regional historical dynamics, and cultural contexts. However, existing LLM-based approaches employ single-model architectures that generate predictions along a singular explicit trajectory, constraining their ability to capture diverse geopolitical nuances across complex regional contexts. To address this limitation, we introduce ThinkTank-ME, a novel Think Tank framework for Middle East event forecasting that emulates collaborative expert analysis in real-world strategic decision-making. To facilitate expert specialization and rigorous evaluation, we construct POLECAT-FOR-ME, a Middle East–focused event forecasting benchmark. Experimental results demonstrate the superiority of multi-expert collaboration in handling complex temporal geopolitical forecasting tasks. The code is available at https://github.com/LuminosityX/ThinkTank-ME.
Most existing content issue detection models are built upon Multimodal Large Language Models (MLLMs). Although such approaches achieve high accuracy, they often suffer from poor timeliness, limited feature richness, and high serving costs: MLLM-based models are typically updated only monthly or quarterly and rely mainly on intrinsic video features (e.g., sampled frames and captions), underutilizing post-hoc user feedback and aggregated video and author-side signals. To address these challenges, we propose a daily updated ranking model for Content Issue Detection based on a Cross-Order Representation Fusion (CORF) architecture. The model supports day-level updates and integrates user behavior features, video- and author-side information, and MLLM-generated scores as inputs. Compared with MLLM-based approaches, our model is significantly smaller and can efficiently score all videos on a daily basis. Experiments on the Adult-Content category show higher recall at the same level of precision, and online A/B tests further demonstrate reduced user exposure to such content issues, contributing to a safer and more positive viewing experience.
With the evolving generative models, generated images are closer to reality, raising concerns about information authenticity and malicious misuse. Invisible watermarks offer a practical approach to detecting and tracing them. However, while image watermarking inevitably introduces quality degradation, most existing methods primarily focus on improving watermark robustness. To address this limitation, we propose AWMA-MoE, a framework that enhances the quality of generated images while preserving strong watermark robustness. Specifically, we design an attention-based adapter that adaptively embeds watermarks with spatially varying strengths across image regions. Building upon this, we introduce an MoE architecture that leverages diverse experts to further improve image quality while retaining watermark robustness. Experiments demonstrate that AWMA-MoE can reduce the distortion of generated images and exhibit competitive watermark performance, thus striking an improved balance for watermarking generated image tasks and better linking post-hoc and in-generation methods.