arXiv subjects

Huaiyu Wan

Publications and source records attributed to Huaiyu Wan.

At least 19 recordsLinked to original sources

arXiv2026

MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents

Credit assignment is challenging in long-horizon agentic reinforcement learning, where supervision often comes only from final rewards. Existing methods refine trajectory-level signals into step-level credits through step grouping or graph-based advantage estimation, but can overlook meaningful intermediate milestones. We propose MileGPO (Milestone Inference with Local Evidence for Graph-Based Policy Optimization), which derives process-level credit from grouped on-policy rollouts through three designs. Milestone Discovery identifies candidate milestones on successful rollouts and recurring traps on failed ones. Reliability-Calibrated Shaping (RCS) weights these candidates by outcome-based confidence, strengthening reliable milestones and traps while down-weighting uncertain ones. Progress-Contrastive Calibration (PCC) further tests whether a candidate reflects local progress and whether its incoming transition outperforms observed alternatives from the same state. MileGPO requires neither auxiliary models nor additional environment interaction. Experiments on ALFWorld and WebShop show state-of-the-art performance and a small in-distribution to out-of-distribution gap on ALFWorld. Ablations and credit diagnostics indicate that reliability weighting, local progress, and same-state branch evidence complement milestone discovery and resolve ambiguous intermediate credit.

cs.LG↗

arXiv2026

Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate

Multi-Agent Debate (MAD) improves the reasoning performance of Large Language Models (LLMs) through multi-round interaction. However, LLMs in MAD are highly susceptible to blind conformity. Existing individual evaluation methods, typically based on confidence or perplexity, fail to reflect the correctness of reasoning and may even exacerbate blind conformity. To address this, we shift the perspective from individual evaluation to group interaction. We define mutual referencing among LLMs as \textbf{Debate Relationships} and recognize that regulating these relationships is the key to mitigating blind conformity. In this paper, we propose a novel framework for \textbf{D}ynamically r\textbf{E}gulating deb\textbf{A}te \textbf{R}elationships (DEAR) from the group perspective. At first, DEAR quantifies consensus and divergence as \textit{group evidence} to capture the debate state. Then, DEAR operates through three stages: 1) What: perceiving group consultation tendency and uncertainty; 2) Who: introducing a Selection RL-Agent to dynamically select reference peers; and 3) How: adopting a Behavior RL-Agent to adaptively adjust generation behaviors. Notably, we formulate the execution of the two RL-Agents as a sequential decision-making process, jointly optimizing via multi-agent reinforcement learning. Extensive experiments demonstrate that DEAR achieves superior performance while significantly reducing token consumption.

cs.MA↗

arXiv2026

Beyond English: Uncovering the Multilingual Gap in Vision-Language-Action Models

Vision-Language-Action models have recently demonstrated promising capabilities in learning generalist robot policies from large-scale multimodal data. However, most existing VLA systems are trained and evaluated primarily with English instructions, leaving their ability to understand and execute instructions in other languages largely unexplored. While the underlying large language models often possess multilingual capabilities, it remains unclear whether these multilingual capabilities transfer to VLAs during training. In this work, we present the first systematic study of multilingual instruction following in VLA models. We first construct multilingual instructions by extending existing benchmarks with translations of their instructions. Using these instructions, we evaluate several representative VLA models across a range of tasks in simulation settings. Our experiments reveal a significant multilingual gap: models trained primarily on English instructions exhibit substantial performance degradation when evaluated on other languages, even when the underlying language backbone is multilingual. We provide several findings and analyses to understand the multilingual gap. Cross-lingual transfer behavior analysis shows that performance drops correlate with both instruction understanding and action execution. Representation analyses suggest that multilingual instruction-caused representation shifts may contribute to the multilingual gap. Motivated by these findings, we further explore strategies to improve multilingual performance in VLAs. We propose a simple yet effective multilingual fine-tuning approach, Multilingual Principal Component Alignment, which leverages Principal Component Analysis to get the principal component subspace and align projected multilingual representations, effectively reducing the multilingual performance gap.

cs.CL↗

arXiv2026

Zoom In, Reason Out: Efficient Far-field Anomaly Detection in Expressway Surveillance Videos via Focused VLM Reasoning Guided by Bayesian Inference

Expressway video anomaly detection is important for traffic safety, but remains challenging across diverse scenes, particularly for far-field vehicles with subtle abnormal motion. Vision-Language Models (VLMs) provide strong semantic reasoning capabilities, yet processing full frames can dilute evidence from distant targets and introduce substantial computational overhead. To address these challenges, we propose VIBES, an asynchronous framework that uses Bayesian inference to guide focused VLM reasoning. Specifically, an online kinematics-guided Bayesian inference module continuously estimates a context-dependent normal-motion distribution from vehicle trajectories and updates its probabilistic boundaries. Deviations from these boundaries produce asynchronous triggers that localize candidate anomalies in time and space. Instead of processing continuous full-frame video, the VLM reasons only over selected frames and localized visual regions associated with the triggers, reducing irrelevant visual content and unnecessary inference. Extensive experiments show that VIBES improves far-field anomaly detection and semantic interpretation while achieving real-time processing efficiency across diverse expressway conditions.

cs.CV↗

arXiv2026

STK-Adapter: Incorporating Evolving Graph and Event Chain for Temporal Knowledge Graph Extrapolation

Temporal Knowledge Graph (TKG) extrapolation aims to predict future events based on historical facts. Recent studies have attempted to enhance TKG extrapolation by integrating TKG's evolving structural representations and textual event chains into Large Language Models (LLMs). Yet, two main challenges limit these approaches: (1) The loss of essential spatial-temporal information due to shallow alignment between TKG's graph evolving structural representation and the LLM's semantic space, and (2) the progressive dilution of the TKG's evolving structural features during LLM fine-tuning. To address these challenges, we propose the Spatial-Temporal Knowledge Adapter (STK-Adapter), which flexibly integrates the evolving graph encoder and the LLM to facilitate TKG reasoning. In STK-Adapter, a Spatial-Temporal MoE is designed to capture spatial structures and temporal patterns inherent in TKGs. An Event-Aware MoE is employed to model intricate temporal semantics dependencies within event chains. In addition, a Cross-Modality Alignment MoE is proposed to facilitate deep cross-modality alignment by TKG-guided attention experts. Extensive experiments on benchmark datasets demonstrate that STK-Adapter significantly outperforms state-of-the-art methods and exhibits strong generalization capabilities in cross-dataset task. The code is available at https://github.com/Zhaoshuyuan0246/STK-Adapter.

cs.IR↗

arXiv2026

SAMoRA: Semantic-Aware Mixture of LoRA Experts for Task-Adaptive Learning

The combination of Mixture-of-Experts (MoE) and Low-Rank Adaptation (LoRA) has shown significant potential for enhancing the multi-task learning capabilities of Large Language Models. However, existing methods face two primary challenges: (1)Imprecise Routing in the current MoE-LoRA method fails to explicitly match input semantics with expert capabilities, leading to weak expert specialization. (2)Uniform weight fusion strategies struggle to provide adaptive update strengths, overlooking the varying complexity of different tasks. To address these limitations, we propose SAMoRA (Semantic-Aware Mixture of LoRA Experts), a novel parameter-efficient fine-tuning framework tailored for task-adaptive learning. Specifically, A Semantic-Aware Router is proposed to explicitly align textual semantics with the most suitable experts for precise routing. A Task-Adaptive Scaling mechanism is designed to regulate expert contributions based on specific task requirements dynamically. In addition, a novel regularization objective is proposed to jointly promote expert specialization and effective scaling. Extensive experiments on multiple multi-task benchmarks demonstrate that SAMoRA significantly outperforms the state-of-the-art methods and holds excellent task generalization capabilities. Code is available at https://github.com/boyan-code/SAMoRA

cs.CL↗

arXiv2026

DiSGMM: A Method for Time-varying Microscopic Weight Completion on Road Networks

Microscopic road-network weights represent fine-grained, time-varying traffic conditions obtained from individual vehicles. An example is travel speeds associated with road segments as vehicles traverse them. These weights support tasks including traffic microsimulation and vehicle routing with reliability guarantees. We study the problem of time-varying microscopic weight completion. During a time slot, the available weights typically cover only some road segments. Weight completion recovers distributions for the weights of every road segment at the current time slot. This problem involves two challenges: (i) contending with two layers of sparsity, where weights are missing at both the network layer (many road segments lack weights) and the segment layer (a segment may have insufficient weights to enable accurate distribution estimation); and (ii) achieving a weight distribution representation that is closed-form and can capture complex conditions flexibly, including heavy tails and multiple clusters. To address these challenges, we propose DiSGMM that combines sparsity-aware embeddings with spatiotemporal modeling to leverage sparse known weights alongside learned segment properties and long-range correlations for distribution estimation. DiSGMM represents distributions of microscopic weights as learnable Gaussian mixture models, providing closed-form distributions capable of capturing complex conditions flexibly. Experiments on two real-world datasets show that DiSGMM can outperform state-of-the-art methods.

cs.LG↗

arXiv2026

Zipage: Maintain High Request Concurrency for LLM Reasoning through Compressed PagedAttention

With reasoning becoming the generative paradigm for large language models (LLMs), the memory bottleneck caused by KV cache during the decoding phase has become a critical factor limiting high-concurrency service. Although existing KV cache eviction methods address the memory issue, most of them are impractical for industrial-grade applications. This paper introduces Compressed PagedAttention, a method that combines token-wise KV cache eviction with PagedAttention. We propose a comprehensive scheduling strategy and support prefix caching and asynchronous compression for Compressed PagedAttention. Based on this, we have developed a high-concurrency LLM inference engine, Zipage. On large-scale mathematical reasoning tasks, Zipage achieves around 95\% of the performance of Full KV inference engines while delivering over 2.1$\times$ speedup.

cs.DC↗

arXiv2026

Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation

Imputing missing values in spatial-temporal traffic data is essential for intelligent transportation systems. Among advanced imputation methods, score-based diffusion models have demonstrated competitive performance. These models generate data by reversing a noising process, using observed values as conditional guidance. However, existing diffusion models typically apply a uniform guidance scale across both spatial and temporal dimensions, which is inadequate for nodes with high missing data rates. Sparse observations provide insufficient conditional guidance, causing the generative process to drift toward the learned prior distribution rather than closely following the conditional observations, resulting in suboptimal imputation performance. To address this, we propose FENCE, a spatial-temporal feedback diffusion guidance method designed to adaptively control guidance scales during imputation. First, FENCE introduces a dynamic feedback mechanism that adjusts the guidance scale based on the posterior likelihood approximations. The guidance scale is increased when generated values diverge from observations and reduced when alignment improves, preventing overcorrection. Second, because alignment to observations varies across nodes and denoising steps, a global guidance scale for all nodes is suboptimal. FENCE computes guidance scales at the cluster level by grouping nodes based on their attention scores, leveraging spatial-temporal correlations to provide more accurate guidance. Experimental results on real-world traffic datasets show that FENCE significantly enhances imputation accuracy.

cs.LG↗

arXiv2026

FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts

Spatial-Temporal Graph (STG) forecasting on large-scale networks has garnered significant attention. However, existing models predominantly focus on short-horizon predictions and suffer from notorious computational costs and memory consumption when scaling to long-horizon predictions and large graphs. Targeting the above challenges, we present FaST, an effective and efficient framework based on heterogeneity-aware Mixture-of-Experts (MoEs) for long-horizon and large-scale STG forecasting, which unlocks one-week-ahead (672 steps at a 15-minute granularity) prediction with thousands of nodes. FaST is underpinned by two key innovations. First, an adaptive graph agent attention mechanism is proposed to alleviate the computational burden inherent in conventional graph convolution and self-attention modules when applied to large-scale graphs. Second, we propose a new parallel MoE module that replaces traditional feed-forward networks with Gated Linear Units (GLUs), enabling an efficient and scalable parallel structure. Extensive experiments on real-world datasets demonstrate that FaST not only delivers superior long-horizon predictive accuracy but also achieves remarkable computational efficiency compared to state-of-the-art baselines. Our source code is available at: https://github.com/yijizhao/FaST.

cs.LG↗

arXiv2026

Rationale-Grounded In-Context Learning for Time Series Reasoning with Multimodal Large Language Models

The underperformance of existing multimodal large language models for time series reasoning lies in the absence of rationale priors that connect temporal observations to their downstream outcomes, which leads models to rely on superficial pattern matching rather than principled reasoning. We therefore propose the rationale-grounded in-context learning for time series reasoning, where rationales work as guiding reasoning units rather than post-hoc explanations, and develop the RationaleTS method. Specifically, we firstly induce label-conditioned rationales, composed of reasoning paths from observable evidence to the potential outcomes. Then, we design the hybrid retrieval by balancing temporal patterns and semantic contexts to retrieve correlated rationale priors for the final in-context inference on new samples. We conduct extensive experiments to demonstrate the effectiveness and efficiency of our proposed RationaleTS on three-domain time series reasoning tasks. We will release our code for reproduction.

cs.AI↗

arXiv2025

RIPCN: A Road Impedance Principal Component Network for Probabilistic Traffic Flow Forecasting

Accurate traffic flow forecasting is crucial for intelligent transportation services such as navigation and ride-hailing. In such applications, uncertainty estimation in forecasting is important because it helps evaluate traffic risk levels, assess forecast reliability, and provide timely warnings. As a result, probabilistic traffic flow forecasting (PTFF) has gained significant attention, as it produces both point forecasts and uncertainty estimates. However, existing PTFF approaches still face two key challenges: (1) how to uncover and model the causes of traffic flow uncertainty for reliable forecasting, and (2) how to capture the spatiotemporal correlations of uncertainty for accurate prediction. To address these challenges, we propose RIPCN, a Road Impedance Principal Component Network that integrates domain-specific transportation theory with spatiotemporal principal component learning for PTFF. RIPCN introduces a dynamic impedance evolution network that captures directional traffic transfer patterns driven by road congestion level and flow variability, revealing the direct causes of uncertainty and enhancing both reliability and interpretability. In addition, a principal component network is designed to forecast the dominant eigenvectors of future flow covariance, enabling the model to capture spatiotemporal uncertainty correlations. This design allows for accurate and efficient uncertainty estimation while also improving point prediction performance. Experimental results on real-world datasets show that our approach outperforms existing probabilistic forecasting methods.

cs.LG↗

arXiv2025

G-KV: Decoding-Time KV Cache Eviction with Global Attention

Recent reasoning large language models (LLMs) excel in complex tasks but encounter significant computational and memory challenges due to long sequence lengths. KV cache compression has emerged as an effective approach to greatly enhance the efficiency of reasoning. However, existing methods often focus on prompt compression or token eviction with local attention score, overlooking the long-term importance of tokens. We propose G-KV, a KV cache eviction method that employs a global scoring mechanism, combining local and historical attention scores to more accurately assess token importance. Additionally, we introduce post-training techniques, including reinforcement learning and distillation, to optimize models for compressed KV cache settings. The code of this paper is available on: https://github.com/microsoft/G-KV.

cs.CL↗

arXiv2025

Think How Your Teammates Think: Active Inference Can Benefit Decentralized Execution

In multi-agent systems, explicit cognition of teammates' decision logic serves as a critical factor in facilitating coordination. Communication (i.e., ``\textit{Tell}'') can assist in the cognitive development process by information dissemination, yet it is inevitably subject to real-world constraints such as noise, latency, and attacks. Therefore, building the understanding of teammates' decisions without communication remains challenging. To address this, we propose a novel non-communication MARL framework that realizes the construction of cognition through local observation-based modeling (i.e., \textit{``Think''}). Our framework enables agents to model teammates' \textbf{active inference} process. At first, the proposed method produces three teammate portraits: perception-belief-action. Specifically, we model the teammate's decision process as follows: 1) Perception: observing environments; 2) Belief: forming beliefs; 3) Action: making decisions. Then, we selectively integrate the belief portrait into the decision process based on the accuracy and relevance of the perception portrait. This enables the selection of cooperative teammates and facilitates effective collaboration. Extensive experiments on the SMAC, SMACv2, MPE, and GRF benchmarks demonstrate the superior performance of our method.

cs.MA↗

arXiv2025

TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model

Vehicle GPS trajectories record how vehicles move over time, storing valuable travel semantics, including movement patterns and travel purposes. Learning travel semantics effectively and efficiently is crucial for real-world applications of trajectory data, which is hindered by two major challenges. First, travel purposes are tied to the functions of the roads and points-of-interest (POIs) involved in a trip. Such information is encoded in textual addresses and descriptions and introduces heavy computational burden to modeling. Second, real-world trajectories often contain redundant points, which harm both computational efficiency and trajectory embedding quality. To address these challenges, we propose TrajMamba, a novel approach for efficient and semantically rich vehicle trajectory learning. TrajMamba introduces a Traj-Mamba Encoder that captures movement patterns by jointly modeling both GPS and road perspectives of trajectories, enabling robust representations of continuous travel behaviors. It also incorporates a Travel Purpose-aware Pre-training procedure to integrate travel purposes into the learned embeddings without introducing extra overhead to embedding calculation. To reduce redundancy in trajectories, TrajMamba features a Knowledge Distillation Pre-training scheme to identify key trajectory points through a learnable mask generator and obtain effective compressed trajectory embeddings. Extensive experiments on two real-world datasets and three downstream tasks show that TrajMamba outperforms state-of-the-art baselines in both efficiency and accuracy.

cs.LG↗

arXiv2025

A Survey of LLM Prompt Datasets: Taxonomy, Linguistic Patterns, and Practical Uses

We compile 129 public LLM prompt datasets with more than 1.22TB and more than 673M instances and organize them into a unified taxonomy. We use seven datasets for detailed analysis and identify lexical, syntactic, and semantic patterns that distinguish prompts from general text. We evaluate these features in prompt filtering, source domain routing, and elicited response quality analysis. A 63 dimensional linguistic feature set that can be extracted on a CPU achieves over 91% of the F1 of GPU sentence embeddings and reduces latency per request by nearly half. In cross dataset routing, sentence embeddings reach 0.74 Macro-F1 and a structural feature subset reaches 0.61, compared with the chance level of 0.20. Structural routing features remain negatively associated with elicited response quality after prompt length is controlled on UltraFeedback. The apparent positive association of type to token ratio becomes much smaller under the same control. These findings support prompt processing pipelines that combine efficient structural features with robust semantic routing. Our catalogue, annotations, and code are given at https://github.com/ymzhang-cs/prompt-dataset-analysis.

cs.LG↗

arXiv2025

A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning

Recent Continual Learning (CL)-based Temporal Knowledge Graph Reasoning (TKGR) methods focus on significantly reducing computational cost and mitigating catastrophic forgetting caused by fine-tuning models with new data. However, existing CL-based TKGR methods still face two key limitations: (1) They usually one-sidedly reorganize individual historical facts, while overlooking the historical context essential for accurately understanding the historical semantics of these facts; (2) They preserve historical knowledge by simply replaying historical facts, while ignoring the potential conflicts between historical and emerging facts. In this paper, we propose a Deep Generative Adaptive Replay (DGAR) method, which can generate and adaptively replay historical entity distribution representations from the whole historical context. To address the first challenge, historical context prompts as sampling units are built to preserve the whole historical context information. To overcome the second challenge, a pre-trained diffusion model is adopted to generate the historical distribution. During the generation process, the common features between the historical and current distributions are enhanced under the guidance of the TKGR model. In addition, a layer-by-layer adaptive replay mechanism is designed to effectively integrate historical and current distributions. Experimental results demonstrate that DGAR significantly outperforms baselines in reasoning and mitigating forgetting.

cs.IR↗

arXiv2025

Enhancing Efficiency and Exploration in Reinforcement Learning for LLMs

Reasoning large language models (LLMs) excel in complex tasks, which has drawn significant attention to reinforcement learning (RL) for LLMs. However, existing approaches allocate an equal number of rollouts to all questions during the RL process, which is inefficient. This inefficiency stems from the fact that training on simple questions yields limited gains, whereas more rollouts are needed for challenging questions to sample correct answers. Furthermore, while RL improves response precision, it limits the model's exploration ability, potentially resulting in a performance cap below that of the base model prior to RL. To address these issues, we propose a mechanism for dynamically allocating rollout budgets based on the difficulty of the problems, enabling more efficient RL training. Additionally, we introduce an adaptive dynamic temperature adjustment strategy to maintain the entropy at a stable level, thereby encouraging sufficient exploration. This enables LLMs to improve response precision while preserving their exploratory ability to uncover potential correct pathways. The code and data is available on: https://github.com/LiaoMengqi/E3-RL4LLMs

cs.LG↗