arXiv subjects

Wenbin Zhang

Publications and source records attributed to Wenbin Zhang.

At least 19 recordsLinked to original sources

arXiv2026

When Does Test-Time Search Help World Action Models? An Empirical Audit of Candidate Selection and Simulator Integrity

World-action models increasingly apply test-time scaling, spending extra inference computation on candidate selection and adaptive allocation to improve control. Such gains are credible only when candidates carry exploitable signal, nominal compute savings become real speedup, and evaluation stays reliable. We audit all three on Cosmos Policy and RoboCasa with a premise-first, pre-registered methodology. On clean data, a learned value selector is 1.66 pp worse than random; a hindsight oracle gains only 1.40 pp, with an episode-level interval spanning zero; and 82% of states show identical candidate outcomes, leaving little selection headroom. Token pruning cuts per-forward FLOPs by up to 47% but yields at most a measured 1.15$\times$ end-to-end speedup on the deployed path (1.25$\times$ on a decode-excluded path), short of a 1.5$\times$ target, and our adaptive-allocation experiment is inconclusive under a power gate. We further identify and fix a silent rendering corruption whose rate rises sharply along compute-heavy planning paths while leaving the official direct-policy result unaffected, and release a detector, fix, and scripts. These findings show why test-time scaling should be audited as a chain of premises (decision signal, realized efficiency, evaluationintegrity) rather than by success rate alone. Conclusions are scoped to the tested configuration.

cs.AI↗

arXiv2026

RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory

Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present \emph{Role-Conditioned Context Routing (RCR)}, an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35\% on average and runtime by 38\% on average relative to Full-Context, while maintaining or improving task performance.

cs.CL↗

arXiv2026

Fairness Definitions in Language Models Explained

Language Models (LMs) have demonstrated exceptional performance across various Natural Language Processing (NLP) tasks. Despite these advancements, LMs can inherit and amplify societal biases related to sensitive attributes such as gender and race, limiting their adoption in real-world applications. Therefore, fairness has been extensively explored in LMs, leading to the proposal of various fairness notions. However, the lack of clear agreement on which fairness definition to apply in specific contexts and the complexity of understanding the distinctions between these definitions can create confusion and impede further progress. To this end, this paper proposes a systematic survey that clarifies the definitions of fairness as they apply to LMs. Specifically, we begin with a brief introduction to LMs and fairness in LMs, followed by a comprehensive, up-to-date overview of existing fairness notions in LMs and the introduction of a novel taxonomy that categorizes these concepts based on their transformer architecture: encoder-only, decoder-only, and encoder-decoder LMs. We further illustrate each definition through experiments, showcasing their practical implications and outcomes. Finally, we discuss current research challenges and open questions, aiming to foster innovative ideas and advance the field. The repository is publicly available online at https://github.com/vanbanTruong/Fairness-in-Large-Language-Models/tree/main/definitions.

cs.CL↗

arXiv2026

LexPilot: Benchmarking the Proactive Interaction Ability of Legal LLMs with Laypeople in Chinese Civil Case Consultation

Large Language Models (LLMs) have greatly assisted legal professionals. However, our empirical studies reveal that, when faced with underspecified consultations from non-experts, current LLMs often produce irresponsible responses with limited utility despite lacking sufficient case details for definitive conclusions. To better assist laypeople, legal LLMs must identify focal points of dispute and proactively elicit crucial context from users. Unfortunately, existing data resources often focus on interaction and response quality, offering limited insight into how proactive interaction affects LLMs' downstream claim-level judgment prediction in actual civil cases. To address this gap, we introduce \textbf{LexPilot}, a benchmark linking proactive legal interaction to claim-level verdict prediction using authentic Chinese civil judgments. Constructed from 432 real-world judgments across 36 common causes of action, LexPilot comprises expert-reviewed pairs of complete and underspecified consultation histories with corresponding court verdicts. Experiments reveal that current LLMs struggle with proactive legal interactions in underspecified consultations, with downstream judgment prediction substantially underperforming that in the fully specified consultation setup (-23.84% on average). Further analysis identifies critical bottlenecks hindering these capabilities. Our work highlights the urgent need for more user-centered legal AI to better serve the public.

cs.AI↗

arXiv2026

Diffusion Reward Models

Reward models underpin the alignment of large language models, yet the dominant designs reduce each prompt--response pair to a point estimate or to a distribution from a fixed parametric family. This is at odds with human preference, which is inherently multimodal: the same response can be reasonably judged in many ways, and no single family covers all of them. To better fit this structure, we introduce DRM, a Diffusion Reward Model that recasts reward modeling as conditional density estimation over $p(\mathbf{r}\mid x,y)$. Conditioned on a frozen LLM encoder, a lightweight Diffusion Transformer denoises Gaussian noise into a reward vector, placing no parametric assumption on the output distribution and naturally representing its multimodal structure. A single architecture handles both multi-attribute regression and pairwise preference data, and at inference $N$ samples form an empirical reward distribution that can be aggregated into a scalar, a variance, or quantiles. Across five benchmarks, DRM matches or surpasses baselines under matched data and backbone, stays competitive with much larger discriminative, distributional, and generative RMs despite its modest training scale, and recovers multimodal reward structure where conventional heads collapse to a point. Uncertainty-aware rejection and lower-confidence-bound (LCB) aggregation further demonstrate that DRM can exploit distributional information beyond a scalar reward to improve reward-model decisions. Downstream RLHF experiments additionally show that using DRM as the training-time reward leads to improved policy performance, directly validating the practical benefit of diffusion-based reward modeling for RLHF training.

cs.LG↗

arXiv2026

Do Agent Benchmarks Do What They Say? An Executable-Contract Audit of Tool-Using Agent Environments

Tool-using agents are entering settings where a wrong action carries real cost, and the benchmarks certifying them grade what each simulated tool call reports having done, assuming the tool did what its interface advertises. The audit taxonomies we survey publish no category for that assumption, and a defect beneath a score is present on every rerun. We treat a tool's advertised surfaces as an executable contract, check the implementation against it, and trace each score's provenance through the task files and evaluator code to the verdicts that derive from state a defective tool should have written. Across 34 audited mutating tools in four benchmarks we confirm seven tool defects and one evaluator property at pinned commits. On injected defects the checker raised no false positive in 25 flags, flagged 2 of 5 negative controls, and missed most: in 29 of 33 scored misses a clause covered the defect but no probe revealed it. The checker's own static half, run alone, flags 14 of 17 confirmed sites, so on these findings the dynamic half confirms and traces rather than discovers. Twelve further AgentDojo tools, with six held-out tools and the seven audited first, complete its 25-tool mutating surface, on which at least 5 tools diverge from their advertised surface as our contracts read it, a rate for AgentDojo alone. No gold trajectory reaches either tau2-bench defect; on 1,120 paths built to isolate the telecom defect, a number fixed by construction, the evaluator rewards a refuel of a suspended line and fails the repaired tool. The clearest case is a clinical benchmark whose tool tells the agent each write executed under a documented no-write design its interface does not disclose; its grader takes that message as evidence, so its action success rate records whether a request carried the expected payload, not whether any record changed.

cs.SE↗

arXiv2026

Towards Extended Active Learning for Modelling Ferroelectric Switching: the Need for 'Gold Standards'

For the purpose of modelling ferroelectric switching in wurtzite-structured materials, four cost-effective density-functional theory (DFT) methods (PBE, PBEsol, r2SCAN, and r2SCAN-rVV10) are considered and compared to various ab initio approaches based on the random-phase approximation (RPA), including RPA with singles corrections (RPAR+S), as well as second-order Møller-Plesset perturbation theory (MP2). The purpose is to determine whether an ab initio approach could act as a 'gold standard' for estimating the reliability of DFT, thus determining an optimal DFT method for use in exhaustive tasks such as the training of machine-learning interatomic potentials (MLIP) for large-scale simulations of materials of arbitrary composition and structure such as Al1-xScxN and Zn1-xMgxO, using AlN, Al0.5Sc0.5N, ZnO, and Zn0.5Mg0.5O as model materials. Applications of active learning (AL) are now common, in which results from MLIP simulations are used to enhance the DFT training data set, but future extended active learning (EAL) methods will also need to systematically assess the DFT methodology against a gold standard. Herein, the variability of the ab initio results is found to exceed that required for a robust gold standard, but the DFT and ab initio approaches appear to converge on RPAR+S and r2SCAN-rVV10 as optimal method choices to initiate EAL. The electron correlation energy is found to be dominated by covalent binding effects associated with the high-electron-density anions involved, but the van der Waals dispersion force is seen to be too significant to ignore in ferroelectric modelling.

cond-mat.mtrl-sci↗

arXiv2026

JudgeSense: A Benchmark for Prompt Sensitivity in LLM-as-a-Judge Systems

Large language models are widely used to judge the output of other language models, yet whether a judge returns the same verdict when the same request is worded differently remains largely unexamined. We study that question across four evaluation tasks and twenty-five judges from six providers. To support the analysis we release JudgeSense, a benchmark of 880 items from human-labelled corpora, each issued under two instructions that differ in wording and not in what they ask, with the complete decision logs. Every score is reported against the judge's own agreement with itself on the identical prompt, so decoding noise is not charged to wording, and the release lets a reader ask the same of any judge not in our roster. Rewording costs agreement on all four tasks, and on two it clears the threshold we declare for a practically meaningful effect; the ordinal task is both the least stable and the one fewest judges are accurate on, and within a single family parameter count does not predict stability. A judge measured inside an agent harness yields a smaller estimate than the same judge reached through a direct API call, because its agreement with itself collapses faster than its agreement across wordings.

cs.CL↗

arXiv2026

FairLMs: A Turnkey Library for Fairness in Language Models

Fairness research on language models involves measuring bias, applying mitigation methods, and examining the evidence on which an evaluation rests. Existing tools offer complementary functionality through different interfaces, so combining them requires reconciling model interfaces, evidence formats, access constraints, and result types before applicability can be checked or methods compared. We introduce \textbf{FairLMs}, a Python library that connects these activities through explicit declarations of model capabilities and input requirements. It provides 33 intrinsic and extrinsic metrics, 14 mitigation components spanning four intervention categories, 14 dataset and scoring-instrument diagnostics, adapters for the three Transformer architectures and supported hosted completion APIs, and benchmark loaders. Declarations are checked before execution and results carry the configuration under which they were obtained, so that compatible components can be combined, methods compared under a common protocol, and workflows extended to new models and datasets. The source code is available at: https://github.com/FairLMs/FairLMs.

cs.LG↗

arXiv2026

Same Patient, Different Order: Action-Level Reliability of Clinical LLM Agents Under Repeated Runs

A clinical agent benchmark can report the same verdict on identical inputs while the agent files a materially different order on each run. Such agents order tests, request medications and place referrals, yet benchmarks typically score one run per task and rarely ask whether identical inputs produce identical actions; MedAgentBench, the benchmark we use, scores a single attempt and says so. To measure this gap we introduce "same-input rerun", which replays a task with every input held fixed and compares the orders rather than the score, with six reliability metrics, and apply it to 1000 MedAgentBench runs across 50 tasks from its five write-capable families, two open-weight models below ten billion parameters quantised to four bits, and two temperatures. The study establishes that action-level divergence exists and can pass unrecorded by the score, not that any rate generalises. Under the 8B model at temperature 0.7, all 43 ordering groups emit a different set of orders across five identical runs, 26 emit the order on some runs and not others, and 28 record a different coded value, dose or analyte. In 22 of those 43 the benchmark reports the same failing verdict for materially different behaviour, as it does for all 10 divergent groups of the 4B model at 0.7. Orders also reach different endpoints across runs, one of which the record server rejects while the agent is told it succeeded. These findings motivate repeated-run evaluation, action-level stability reporting and execution-faithful environment feedback in clinical-agent benchmarks.

cs.CL↗

arXiv2026

When Should a Robot Think? Resource-Aware Reasoning via Reinforcement Learning for Embodied Robotic Decision-Making

Embodied robotic systems increasingly rely on large language model (LLM)-based agents to support high-level reasoning, planning, and decision-making during interactions with the environment. However, invoking LLM reasoning introduces substantial computational latency and resource overhead, which can interrupt action execution and reduce system reliability. Excessive reasoning may delay actions, while insufficient reasoning often leads to incorrect decisions and task failures. This raises a fundamental question for embodied agents: when should the agent reason, and when should it act? In this work, we propose RARRL (Resource-Aware Reasoning via Reinforcement Learning), a hierarchical framework for resource-aware orchestration of embodied agents. Rather than learning low-level control policies, RARRL learns a high-level orchestration policy that operates at the agent's decision-making layer. This policy enables the agent to adaptively determine whether to invoke reasoning, which reasoning role to employ, and how much computational budget to allocate based on current observations, execution history, and remaining resources. Extensive experiments, including evaluations with empirical latency profiles derived from the ALFRED benchmark, show that RARRL consistently improves task success rates while reducing execution latency and enhancing robustness compared with fixed or heuristic reasoning strategies. These results demonstrate that adaptive reasoning control is essential for building reliable and efficient embodied robotic agents.

cs.RO↗

arXiv2026

C2C-Explorer: An Exploration Framework for Chip-to-Chip Interconnect Architectures in LLM Cloud Computing Systems

The scaling-up of large language models (LLMs) necessitates computing systems to have multi-processor-chip architectures, elevating the importance of chip-to-chip (C2C) communication. However, designing efficient C2C hardware architectures for LLM workloads faces three key challenges: generating realistic LLM-specific C2C traffic, accurately simulating hardware-level communication at scale, and efficiently exploring the exponentially large C2C design space. We propose C2C-Explorer, an adaptive Bayesian DSE framework that integrates a LLM-workload-driven traffic generator, a scalable interconnect simulator (switch/full-mesh, up to 512 chips), and a metric-guided evaluator into a workload-to-hardware optimization pipeline, enabling systematic C2C architectural co-design under realistic LLM workloads. Validated against FPGA-based C2C prototypes, the C2C simulator achieves 2.46-8.23% end-to-end timing error across diverse traffic patterns. Its hybrid cycle and event model further accelerates large-scale simulation by up to 7.8$\times$ over a pure cycle-accurate baseline. Applied to a 32-XPU DeepSeek-R1-671B inference workload, C2C-Explorer identifies configurations that improve goodput by 44.1% and reduce memory by 98.4%. C2C-Explorer is open-source and available at https://github.com/Selinaee/C2C-Explorer.

cs.DC↗

arXiv2026

Structured Agent Distillation for Large Language Model

Large language models (LLMs) exhibit strong capabilities as decision-making agents by interleaving reasoning and actions, as seen in ReAct-style frameworks. Yet, their practical deployment is constrained by high inference costs and large model sizes. We propose Structured Agent Distillation, a framework that compresses large LLM-based agents into smaller student models while preserving both reasoning fidelity and action consistency. Unlike standard token-level distillation, our method segments trajectories into [REASON] and [ACT] spans, applying segment-specific losses to align each component with the teacher's behavior. This structure-aware supervision enables compact agents to better replicate the teacher's decision process. Experiments on ALFWorld, HotPotQA-ReAct, and WebShop show that our approach consistently outperforms token-level and imitation learning baselines, achieving significant compression with minimal performance drop. Scaling and ablation results further highlight the importance of span-level alignment for efficient and deployable agents.

cs.LG↗

arXiv2026

A Technical Policy Blueprint for Trustworthy Decentralized AI

Decentralized AI systems, such as federated learning, can play a critical role in further unlocking AI asset marketplaces (e.g., healthcare data marketplaces) thanks to increased asset privacy protection. Unlocking this big potential necessitates governance mechanisms that are transparent, scalable, and verifiable. However current governance approaches rely on bespoke, infrastructure-specific policies that hinder asset interoperability and trust among systems. We are proposing a Technical Policy Blueprint that encodes governance requirements as policy-as-code objects and separates asset policy verification from asset policy enforcement. In this architecture the Policy Engine verifies evidence (e.g., identities, signatures, payments, trusted-hardware attestations) and issues capability packages. Asset Guardians (e.g. data guardians, model guardians, computation guardians, etc.) enforce access or execution solely based on these capability packages. This core concept of decoupling policy processing from capabilities enables governance to evolve without reconfiguring AI infrastructure, thus creating an approach that is transparent, auditable, and resilient to change.

cs.CY↗

arXiv2026

Elucidating Representation Degradation Problem in Diffusion Model Training

Diffusion models have achieved remarkable success, yet their training remains inefficient due to a severe optimization bottleneck, which we term Representation Degradation. As noise levels increase, the outputs of the trained model exhibit progressive structural distortion, which can destabilize training and impair generation quality. Our analysis suggests that this instability is driven by mismatched target recoverability, which is associated with Neural Tangent Kernel (NTK) spectral weakening and effective low-rank behavior. To address this, we propose Elucidated Representation Diffusion (ERD), a plug-and-play framework that dynamically reallocates optimization effort according to effective recoverability. By stabilizing representation learning without external supervision, ERD accelerates convergence and achieves strong empirical performance across diffusion backbones.

cs.LG↗

arXiv2026

Geospatial foundation-model embeddings improve population estimation unevenly across space and scale

Reliable subnational population estimates are essential for applications, yet remain difficult where censuses are sparse, outdated or spatially coarse. Existing population-mapping workflows rely on hand-built geospatial covariates, such as settlement extent, night-time lights, and environmental conditions, which must be assembled and harmonised across scales and geographies. Geospatial foundation models offer an alternative by learning reusable representations of place from more multifaceted and heterogeneous data sources. Here, we benchmark Population Dynamics Foundation Model (PDFM) embeddings against the harmonised geospatial covariates for subnational population estimation in Brazil, Nigeria and the United States. Under geographically structured validation, PDFM increased predictive fit by a median of 20.1% (IQR: 10.0-33.2%, across country-model comparisons) reduction in unexplained variance, and reduced Kullback-Leibler divergence by 23.2% (9.2-26.2%). However, these gains were uneven. PDFM was most advantageous where the geospatial covariates weakly characterised settlement context, such as larger and less-developed subnational areas. Moreover, PDFM performance was scale-coupled with embeddings providing less flexible transfer across spatial aggregations than geospatial covariates. These findings showed that geospatial foundation-model representations of place can improve population estimation in data poor settings, but their benefits break down predictably under spatial scale mismatch, revealing a fundamental limitation of current geospatial AI.

cs.LG↗

arXiv2026

Learning to Conceal Risk: Controllable Multi-turn Red Teaming for LLMs in the Financial Domain

Large Language Models (LLMs) are increasingly deployed in finance, where unsafe behavior can lead to serious regulatory risks. However, most red-teaming research focuses on overtly harmful content and overlooks attacks that appear legitimate on the surface yet induce regulatory-violating responses. We address this gap by introducing a controllable black-box multi-turn risk-concealed red-teaming framework (CoRT) that progressively conceals surface-level risk while exploiting regulatory-violating behaviors. CoRT contains two key components: (i) a Risk Concealment Attacker (RCA) that generates multi-turn prompts via iterative refinement, and (ii) a Risk Concealment Controller (RCC) that predicts a turn-level Risk Concealment Score (RCS) to steer RCA's follow-up style. We also built a domain-specific benchmark, FinRisk-Bench, with 522 instructions spanning six financial risk categories. Experiments on nine widely used LLMs show that CoRT (RCA) achieves 93.19% average attack success rate (ASR), and CoRT (RCA+RCC) further improves the average ASR to 95.00%. Our code and FinRisk-Bench are available at https://github.com/gcheng128/CoRT.

cs.CL↗

arXiv2026

Towards Fair Machine Learning Software: Understanding and Addressing Model Bias Through Counterfactual Thinking

The increasing use of Machine Learning (ML) software can lead to unfair and unethical decisions, thus fairness bugs in software are becoming a growing concern. Addressing these fairness bugs often involves sacrificing ML performance, such as accuracy. To address this issue, we present a novel counterfactual approach that uses counterfactual thinking to tackle the root causes of bias in ML software. In addition, our approach combines models optimized for both performance and fairness, resulting in an optimal solution in both aspects. We conducted a thorough evaluation of our approach on 10 benchmark tasks using a combination of 5 performance metrics, 3 fairness metrics, and 15 measurement scenarios, all applied to 8 real-world datasets. The conducted extensive evaluations show that the proposed method significantly improves the fairness of ML software while maintaining competitive performance, outperforming state-of-the-art solutions in 84.6% of overall cases based on a recent benchmarking tool.

cs.SE↗