arXiv subjects

Wenbo Chen

Publications and source records attributed to Wenbo Chen.

At least 19 recordsLinked to original sources

arXiv2026

BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure

LM-agent benchmarks increasingly function as interactive evaluation infrastructure. Agents observe state, call tools, modify workspaces, submit artifacts, and receive rewards from outcome procedures. This interactivity makes evaluations vulnerable to reward hacking: an agent improves its measured score by exploiting the reward-relevant trajectory instead of solving the intended task. Existing defenses rely largely on task-specific patches, prompt instructions, or post-hoc detectors. They do not provide reusable evidence that a concrete run remained within its intended evaluation boundary. This paper presents BenchShield, a model-backed instrumentation layer for reward integrity in LLM-agent evaluation. BenchShield grounds detection in a finite lifecycle model of an evaluation's reward-relevant events. Within the benchmark infrastructure, two complementary analyses operate over this model. A static, phase-aware taint analysis exposes reward-hacking paths before a run. Its runtime counterpart uses infrastructure-side evidence to attribute concrete agent use and emit evidence-backed claims. We construct BenchShield Trajectories, a human-labeled corpus of 456 adjudicated trajectories from more than 31,000 public agent runs across three benchmarks. Compared with an agentic hackability scanner baseline on the same tasks and model, BenchShield improves full-chain recall from 23-94% to 77-100%, same-vector coverage from 16-56% to 43-78%, and reduces per-task cost by up to 65%. Its runtime analysis achieves 96% accuracy in detecting reward hacking from infrastructure-side evidence.

cs.CR↗

arXiv2026

MARS: Margin-Adversarial Risk-controlled Stopping for Parallel LLM Test-time Scaling

Parallel test-time scaling samples many reasoning traces and majority-votes their answers, improving LLM accuracy but requiring traces to run to completion, incurring substantial computational overhead. We observe that probing partial traces at intermediate checkpoints can extract current answers without disrupting generation, revealing an evolving aggregate vote. Based on this observation, we introduce MARS, a margin-adversarial stopping rule that estimates which active traces are likely to change their answers and stops once the leader remains safe under a conservative bound on future vote movement. The rule separates two sources of uncertainty. It learns the trace-level switch probabilities that determine how much of the current margin is likely to be retained, while handling the harder question of where switching traces land through an adversarial bound calibrated from warmup traces. With true switch probabilities, MARS guarantees with high probability that the early-stopped answer matches the full-budget vote. In practice, a five-feature logistic model closely matches oracle switching behavior. Across three reasoning models and three competition-math benchmarks, MARS saves 25-47% of self-consistency tokens and 14-29% on top of DeepConf Online, a strong confidence-weighted baseline that already filters and truncates weak traces, while matching the accuracy of the corresponding full-budget baselines.

cs.AI↗

arXiv2026

Structured Prompt Optimization Meets Reinforcement Learning for Global and Local Interpretability over Complex Text

LLMs have advanced text classification, yet existing paradigms face a trade-off: supervised (label only) fine-tuning is scalable but offers limited reasoning on complex text and lacks broader model transparency, while discrete prompt optimization offers human-readable instructions but struggles with performance and scalability. We introduce eXTC (eXplainable Text Classifier) with three progressive stages: (1) learning a Standard Operating Procedure (SOP, or rulebook) in natural language via a new Structured Prompt Optimization algorithm; (2) SOP-grounded reasoning distillation from a large teacher LLM into a compact LM; and (3) expanding reasoning capabilities beyond the initial SOP via reinforcement learning. This design enables eXTC to provide (i) fast inference via a compact LM, with (ii) inference-time local reasoning traces, alongside a global, modular explanation of its learned domain rules, while (iii) significantly outperforming existing paradigms across diverse benchmarks in both classification performance and explanation quality, with stage-by-stage gains.

cs.CL↗

arXiv2026

HCSG: Human-Centric Semantic-Geometric Reasoning for Vision-Language Navigation

VLN has achieved remarkable progress by scaling data and model capacity. However, the assumption of a static environment breaks down in real-world indoor scenarios, where robots inevitably encounter dynamic pedestrians. Existing human-aware approaches typically treat humans merely as moving obstacles based on implicit visual cues, lacking the explicit reasoning required to interpret human intentions or maintain social norms. To address this, we propose HCSG, the first human-centric framework for VLN. This framework provides a robust foundation for safe, socially intelligent navigation in dynamic human-robot environments that shifts the paradigm from passive collision avoidance to active human behavior understanding. Specifically, HCSG introduces a unified Human Understanding Module that synergizes two key capabilities: (i) geometric forecasting, which predicts human pose and trajectory to anticipate future motion dynamics; and (ii) semantic interpretation, which leverages a Vision-Language Model (VLM) to generate natural language descriptions of human actions and intentions. These semantic-geometric representations are fused into the agent's topological map for instruction-conditioned planning. Furthermore, a social distance loss is introduced to enforce socially compliant interaction distances. Extensive experiments on the HA-VLNCE benchmark demonstrate that HCSG significantly outperforms state-of-the-art methods, achieving a 14% improvement in Success Rate and a 34% reduction in Collision Rate. Our project can be seen at https://haoxuanxu1024.github.io/HCSG/.

cs.RO↗

arXiv2026

Rethinking Evaluation for LLM Hallucination Detection: A Desiderata, A New RAG-based Benchmark, New Insights

Hallucination, broadly referring to unfaithful, fabricated, or inconsistent content generated by LLMs, has wide-ranging implications. Therefore, a large body of effort has been devoted to detecting LLM hallucinations, as well as designing benchmark datasets for evaluating these detectors. In this work, we first establish a desiderata of properties for hallucination detection benchmarks (HDBs) to exhibit for effective evaluation. A critical look at existing HDBs through the lens of our desiderata reveals that none of them exhibits all the properties. We identify two largest gaps: (1) RAG-based grounded benchmarks with long context are severely lacking (partly because length impedes human annotation); and (2) Existing benchmarks do not make available realistic label noise for stress-testing detectors although real-world use-cases often grapple with label noise due to human or automated/weak annotation. To close these gaps, we build and open-source a new RAG-based HDB called T RIVIA+ that underwent a rigorous human annotation process. Notably, our benchmark exhibits all desirable properties including (1) T RIVIA+ contains samples with the longest context in the literature; and (2) we design and share four sets of noisy labels with different, both sample-dependent and sampleindependent, noise schemes. Finally, we perform experiments on RAG-based HDBs, including our T RIVIA+, using popular SOTA detectors that reveal new insights: (i) ample room remains for current detectors to reach the performance ceiling on RAG-based HDBs, (ii) the basic LLM-as-a-Judge baseline performs competitively, and (iii) label noise hinders detection performance. We expect that our findings, along with our proposed benchmark 1 , will motivate and foster needed research on hallucination detection for RAG-based tasks.

cs.AI↗

arXiv2026

ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces

Large language model (LLM) agents are increasingly deployed to automate productivity tasks (e.g., email, scheduling, document management), but evaluating them on live services is risky due to potentially irreversible changes. Existing benchmarks rely on simplified environments and fail to capture realistic, stateful, multi-service workflows. We introduce ClawsBench, a benchmark for evaluating and improving LLM agents in realistic productivity settings. It includes five high-fidelity mock services (Gmail, Slack, Google Calendar, Google Docs, Google Drive) with full state management and deterministic snapshot/restore, along with 44 structured tasks covering single-service, cross-service, and safety-critical scenarios. We decompose agent scaffolding into two independent levers (domain skills that inject API knowledge via progressive disclosure, and a meta prompt that coordinates behavior across services) and vary both to measure their separate and combined effects. Experiments across 6 models, 4 agent harnesses, and 33 conditions show that with full scaffolding, agents achieve task success rates of 39-64% but exhibit unsafe action rates of 7-33%. On OpenClaw, the top five models fall within a 10 percentage-point band on task success (53-63%), with unsafe action rates from 7% to 23% and no consistent ordering between the two metrics. We identify eight recurring patterns of unsafe behavior, including multi-step sandbox escalation and silent contract modification. We release the trajectories and future dataset at https://clawsbench.com.

cs.AI↗

arXiv2026

P$^{3}$Nav: End-to-End Perception, Prediction and Planning for Vision-and-Language Navigation

In Vision-and-Language Navigation (VLN), an agent is required to plan a path to the target specified by the language instruction, using its visual observations. Consequently, prevailing VLN methods primarily focus on building powerful planners through visual-textual alignment. However, these approaches often bypass the imperative of comprehensive scene understanding prior to planning, leaving the agent with insufficient perception or prediction capabilities. Thus, we propose P$^{3}$Nav, a novel end-to-end framework integrating perception, prediction, and planning in a unified pipeline to strengthen the VLN agent's scene understanding and boost navigation success. Specifically, P$^{3}$Nav augments perception by extracting complementary cues from object-level and map-level perspectives. Subsequently, our P$^{3}$Nav predicts waypoints to model the agent's potential future states, endowing the agent with intrinsic awareness of candidate positions during navigation. Conditioned on these future waypoints, P$^{3}$Nav further forecasts semantic map cues, enabling proactive planning and reducing the strict reliance on purely historical context. Integrating these perceptual and predictive cues, a holistic planning module finally carries out the VLN tasks. Extensive experiments demonstrate that our P$^{3}$Nav achieves new state-of-the-art performance on the REVERIE, R2R-CE, and RxR-CE benchmarks.

cs.RO↗

arXiv2026

Enhancing Vision-Language Navigation with Multimodal Event Knowledge from Real-World Indoor Tour Videos

Vision-Language Navigation (VLN) agents often struggle with long-horizon reasoning in unseen environments, particularly when facing ambiguous, coarse-grained instructions. While recent advances use knowledge graph to enhance reasoning, the potential of multimodal event knowledge inspired by human episodic memory remains underexplored. In this work, we propose an event-centric knowledge enhancement strategy for automated process knowledge mining and feature fusion to solve coarse-grained instruction and long-horizon reasoning in VLN task. First, we construct YE-KG, the first large-scale multimodal spatiotemporal knowledge graph, with over 86k nodes and 83k edges, derived from real-world indoor videos. By leveraging multimodal large language models (i.e., LLaVa, GPT4), we extract unstructured video streams into structured semantic-action-effect events to serve as explicit episodic memory. Second, we introduce STE-VLN, which integrates the above graph into VLN models via a Coarse-to-Fine Hierarchical Retrieval mechanism. This allows agents to retrieve causal event sequences and dynamically fuse them with egocentric visual observations. Experiments on REVERIE, R2R, and R2R-CE benchmarks demonstrate the efficiency of our event-centric strategy, outperforming state-of-the-art approaches across diverse action spaces. Our data and code are available on the project website https://sites.google.com/view/y-event-kg/.

cs.RO↗

arXiv2026

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

cs.AI↗

arXiv2026

Latent Spherical Flow Policy for Reinforcement Learning with Combinatorial Actions

Reinforcement learning (RL) with combinatorial action spaces remains challenging because feasible action sets are exponentially large and governed by complex feasibility constraints, making direct policy parameterization impractical. Existing approaches embed task-specific value functions into constrained optimization programs or learn deterministic structured policies, sacrificing generality and policy expressiveness. We propose a solver-induced \emph{latent spherical flow policy} that brings the expressiveness of modern generative policies to combinatorial RL while guaranteeing feasibility by design. Our method, LSFlow, learns a \emph{stochastic} policy in a compact continuous latent space via spherical flow matching, and delegates feasibility to a combinatorial optimization solver that maps each latent sample to a valid structured action. To improve efficiency, we train the value network directly in the latent space, avoiding repeated solver calls during policy optimization. To address the piecewise-constant and discontinuous value landscape induced by solver-based action selection, we introduce a smoothed Bellman operator that yields stable, well-defined learning targets. Empirically, our approach outperforms state-of-the-art baselines by an average of 20.6\% across a range of challenging combinatorial RL tasks.

cs.LG↗

arXiv2025

BRAINCELL-AID: An Agentic AI Created Brain Cell Type Resource for Community Annotation

Single-cell RNA sequencing has transformed our ability to identify diverse cell types and their transcriptomic signatures. However, annotating these signatures-especially those involving poorly characterized genes-remains a major challenge. Traditional methods, such as Gene Set Enrichment Analysis (GSEA), depend on well-curated annotations and often perform poorly in these contexts. Large Language Models (LLMs) offer a promising alternative but struggle to represent complex biological knowledge within structured ontologies. To address this, we present BRAINCELL-AID (BRAINCELL-AID: https://biodataai.uth.edu/BRAINCELL-AID), a novel multi-agent AI system that integrates free-text descriptions with ontology labels to enable more accurate and robust gene set annotation. By incorporating retrieval-augmented generation (RAG), we developed a robust agentic workflow that refines predictions using relevant PubMed literature, reducing hallucinations and enhancing interpretability. Using this workflow, we achieved correct annotations for 77% of mouse gene sets among their top predictions. Applying this approach, we annotated 5,322 brain cell clusters from the comprehensive mouse brain cell atlas generated by the BRAIN Initiative Cell Census Network, enabling novel insights into brain cell function by identifying region-specific gene co-expression patterns and inferring functional roles of gene ensembles. BRAINCELL-AID also identifies Basal Ganglia-related cell types with neurologically meaningful descriptions. Hence, we create a valuable resource to support community-driven cell type annotation.

cs.AI↗

arXiv2025

MCOP: Multi-UAV Collaborative Occupancy Prediction

Unmanned Aerial Vehicle (UAV) swarm systems necessitate efficient collaborative perception mechanisms for diverse operational scenarios. Current Bird's Eye View (BEV)-based approaches exhibit two main limitations: bounding-box representations fail to capture complete semantic and geometric information of the scene, and their performance significantly degrades when encountering undefined or occluded objects. To address these limitations, we propose a novel multi-UAV collaborative occupancy prediction framework. Our framework effectively preserves 3D spatial structures and semantics through integrating a Spatial-Aware Feature Encoder and Cross-Agent Feature Integration. To enhance efficiency, we further introduce Altitude-Aware Feature Reduction to compactly represent scene information, along with a Dual-Mask Perceptual Guidance mechanism to adaptively select features and reduce communication overhead. Due to the absence of suitable benchmark datasets, we extend three datasets for evaluation: two virtual datasets (Air-to-Pred-Occ and UAV3D-Occ) and one real-world dataset (GauUScene-Occ). Experiments results demonstrate that our method achieves state-of-the-art accuracy, significantly outperforming existing collaborative methods while reducing communication overhead to only a fraction of previous approaches.

cs.CV↗

arXiv2025

Boosting Column Generation with Graph Neural Networks for Joint Rider Trip Planning and Crew Shift Scheduling

Optimizing service schedules is pivotal to the reliable, efficient, and inclusive on-demand mobility. This pressing challenge is further exacerbated by the increasing needs of an aging population, the oversubscription of existing services, and the lack of effective solution methods. This study addresses the intricacies of service scheduling, by jointly optimizing rider trip planning and crew scheduling for a complex dynamic mobility service. The resulting optimization problems are extremely challenging computationally for state-of-the-art methods. To address this fundamental gap, this paper introduces the Joint Rider Trip Planning and Crew Shift Scheduling Problem (JRTPCSSP) and a novel solution method, called Attention and Gated GNN-Informed Column Generation (AGGNNI-CG), that hybridizes column generation and machine learning to obtain near-optimal solutions to the JRTPCSSP with real-life constraints of the application. The key idea of the machine-learning component is to dramatically reduce the number of paths to explore in the pricing problem, accelerating the most time-consuming component of the column generation. The machine learning component is a graph neural network with an attention mechanism and a gated architecture, which is particularly suited to cater for the different input sizes coming from daily operations. AGGNNI-CG has been applied to a challenging, real-world dataset from the Paratransit system of Chatham County in Georgia. It produces substantial improvements compared to the baseline column generation approach, which typically cannot produce high-quality feasible solutions in reasonable time on large-scale complex instances. AGGNNI-CG also produces significant improvements in service quality compared to the existing system.

math.OC↗

arXiv2024

GPTON: Generative Pre-trained Transformers enhanced with Ontology Narration for accurate annotation of biological data

By leveraging GPT-4 for ontology narration, we developed GPTON to infuse structured knowledge into LLMs through verbalized ontology terms, achieving accurate text and ontology annotations for over 68% of gene sets in the top five predictions. Manual evaluations confirm GPTON's robustness, highlighting its potential to harness LLMs and structured knowledge to significantly advance biomedical research beyond gene set annotation.

q-bio.QM↗

arXiv2024

Compact Optimality Verification for Optimization Proxies

Recent years have witnessed increasing interest in optimization proxies, i.e., machine learning models that approximate the input-output mapping of parametric optimization problems and return near-optimal feasible solutions. Following recent work by (Nellikkath & Chatzivasileiadis, 2021), this paper reconsiders the optimality verification problem for optimization proxies, i.e., the determination of the worst-case optimality gap over the instance distribution. The paper proposes a compact formulation for optimality verification and a gradient-based primal heuristic that brings substantial computational benefits to the original formulation. The compact formulation is also more general and applies to non-convex optimization problems. The benefits of the compact formulation are demonstrated on large-scale DC Optimal Power Flow and knapsack problems.

math.OC↗

arXiv2024

Optimization-based Learning for Dynamic Load Planning in Trucking Service Networks

The load planning problem is a critical challenge in service network design for parcel carriers: it decides how many trailers to assign for dispatch over time between pairs of terminals. Another key challenge is to determine a flow plan, which specifies how parcel volumes are assigned to planned loads. This paper considers the Outbound Load Planning Problem (OLPP) that considers flow and load planning challenges jointly in order to adjust loads and flows as the demand forecast changes over time before the day of operations in a terminal. The paper aims at developing a decision-support tool to inform planners making these decisions at terminals across the network. The paper formulates the OLPP as a mixed-integer programming model and shows that it admits a large number of symmetries in a network where each commodity can be routed through primary and alternate terminals. As a result, an optimization solver may return fundamentally different solutions to closely related problems, confusing planners and reducing trust in optimization. To remedy this limitation, this paper proposes a lexicographical optimization approach that eliminates those symmetries by generating optimal solutions staying close to a reference plan. Moreover, this paper designs an optimization proxy that addresses the computational challenges of the optimization model. The optimization proxy combines a machine-learning model and a repair procedure to find near-optimal solutions that satisfy real-time constraints imposed by planners in the loop. An extensive computational study on industrial instances shows that the optimization proxy is orders of magnitude faster for generating solutions that are consistent with each other. The proposed approach also demonstrates the benefits of the OLPP for load consolidation and the significant savings obtained from combining machine learning and optimization.

cs.AI↗

arXiv2023

Real-Time Risk Analysis with Optimization Proxies

The increasing penetration of renewable generation and distributed energy resources requires new operating practices for power systems, wherein risk is explicitly quantified and managed. However, traditional risk-assessment frameworks are not fast enough for real-time operations, because they require numerous simulations, each of which requires solving multiple economic dispatch problems sequentially. The paper addresses this computational challenge by proposing proxy-based risk assessment, wherein optimization proxies are trained to learn the input-to-output mapping of an economic dispatch optimization solver. Once trained, the proxies make predictions in milliseconds, thereby enabling real-time risk assessment. The paper leverages self-supervised learning and end-to-end-feasible architecture to achieve high-quality sequential predictions. Numerical experiments on large systems demonstrate the scalability and accuracy of the proposed approach.

math.OC↗

arXiv2023

End-to-End Feasible Optimization Proxies for Large-Scale Economic Dispatch

The paper proposes a novel End-to-End Learning and Repair (E2ELR) architecture for training optimization proxies for economic dispatch problems. E2ELR combines deep neural networks with closed-form, differentiable repair layers, thereby integrating learning and feasibility in an end-to-end fashion. E2ELR is also trained with self-supervised learning, removing the need for labeled data and the solving of numerous optimization problems offline. E2ELR is evaluated on industry-size power grids with tens of thousands of buses using an economic dispatch that co-optimizes energy and reserves. The results demonstrate that the self-supervised E2ELR achieves state-of-the-art performance, with optimality gaps that outperform other baselines by at least an order of magnitude.

math.OC↗