arXiv subjects

Canyu Chen

Publications and source records attributed to Canyu Chen.

At least 19 recordsLinked to original sources

arXiv2026

Agentic Commerce World: An Auditable and Verifiable Environment for Vibe Commerce

In vibe coding, people describe software in natural language and delegate implementation to AI agents. By analogy, vibe commerce allows people to express buying or selling goals in natural language and delegate the corresponding tasks to agents. Commerce, however, requires independently controlled Buyer and Merchant agents to interact in a shared market while preserving their private objectives and distinct authority. We introduce Agentic Commerce World (ACWorld), an environment for evaluating such agents across ongoing transactions. Through its Vibe Commerce Protocol (VCP), ACWorld validates agent actions before updating shared transaction state and records the resulting interactions, making agent behavior auditable and evaluation reproducible. The ACWorld Benchmark contains a 200-task capability-coverage track and a 60-task large-catalog track that searches 785,022 transactable listings. Across ten models, mean scores range from 65.9% to 85.6% and from 56.1% to 91.4%, respectively. Our analysis shows that process-level evidence is necessary: final state alone can miss evaluated errors, incomplete trajectories still retain useful process signals, and large-catalog tasks expose bottlenecks across stages.

cs.AI↗

arXiv2026

Teaching Vision-Language-Action Models What to See and Where to Look

Vision-Language-Action (VLA) models have emerged as a promising paradigm for end-to-end autonomous driving. However, existing VLAs' training relies heavily on text-centric visual question answering and chain-of-thought reasoning data, which emphasizes linguistic reasoning rather than action-grounded planning. As a result, the learned representations capture semantic knowledge but lack spatial dependencies crucial for reliable trajectory prediction. We propose DriveTeach-VLA, a framework that explicitly teaches VLAs what to see and where to look. Driving-aware Vision Distillation (DVD) injects driving-specific perceptual priors into the vision encoder, while 2D Trajectory-Guided Prompts (2D-TGP) provide spatial conditioning aligned with feasible driving trajectories. Together, they form a vision-guided learning pipeline: what to see (DVD pretraining) - where to look (TGP-guided SFT) - how to act (TGP-guided GRPO). DriveTeach-VLA achieves the state-of-the-art performance on NAVSIM and nuScenes. Our code is available at: https://github.com/ShivaTeam/DriveTeach-VLA.

cs.CV↗

arXiv2026

CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planning

End-to-end autonomous driving planners are commonly trained by imitating a single logged trajectory, yet evaluated by rule-based planning metrics that measure safety, feasibility, progress, and comfort. This creates a training--evaluation mismatch: trajectories close to the logged path may violate planning rules, while alternatives farther from the demonstration can remain valid and high-scoring. The mismatch is especially limiting for proposal-selection planners, whose performance depends on candidate-set coverage and scorer ranking quality. We propose CLOVER, a Closed-LOop Value Estimation and Ranking framework for end-to-end autonomous driving planning. CLOVER follows a lightweight generator--scorer formulation: a generator produces diverse candidate trajectories, and a scorer predicts planning-metric sub-scores to rank them at inference time. To expand proposal support beyond single-trajectory imitation, CLOVER constructs evaluator-filtered pseudo-expert trajectories and trains the generator with set-level coverage supervision. It then performs conservative closed-loop self-distillation: the scorer is fitted to true evaluator sub-scores on generated proposals, while the generator is refined toward teacher-selected top-$k$ and vector-Pareto targets with stability regularization. We analyze when an imperfect scorer can improve the generator, showing that scorer-mediated refinement is reliable when scorer-selected targets are enriched under the true evaluator and updates remain conservative. On NAVSIM, CLOVER achieves 94.5 PDMS and 90.4 EPDMS, establishing a new state of the art. On the more challenging NavHard split, it obtains 48.3 EPDMS, matching the strongest reported result. On supplementary nuScenes open-loop evaluation, CLOVER achieves the lowest L2 error and collision rate among compared methods. Code data will be released at https://github.com/WilliamXuanYu/CLOVER.

cs.RO↗

arXiv2026

SURGE: Surrogate Gradient Adaptation in Binary Neural Networks

The training of Binary Neural Networks (BNNs) is fundamentally based on gradient approximation for non-differentiable binarization operations (e.g., sign function). However, prevailing methods including the Straight-Through Estimator (STE) and its improved variants, rely on hand-crafted designs that suffer from gradient mismatch problem and information loss induced by fixed-range gradient clipping. To address this, we propose SURrogate GradiEnt Adaptation (SURGE), a novel learnable gradient compensation framework with theoretical grounding. SURGE mitigates gradient mismatch through auxiliary backpropagation. Specifically, we design a Dual-Path Gradient Compensator (DPGC) that constructs a parallel full-precision auxiliary branch for each binarized layer, decoupling gradient flow via output decomposition during backpropagation. DPGC enables bias-reduced gradient estimation by leveraging the full-precision branch to estimate components beyond STE's first-order approximation. To further enhance training stability, we introduce an Adaptive Gradient Scaler (AGS) based on an optimal scale factor to dynamically balance inter-branch gradient contributions via norm-based scaling. Experiments on image classification, object detection, and language understanding tasks demonstrate that SURGE performs best over state-of-the-art methods.

cs.LG↗

arXiv2026

SilLang: Improving Gait Recognition with Silhouette Language Encoding

Gait silhouettes, which can be encoded into binary gait codes, are widely adopted to representing motion patterns of pedestrian. Recent approaches commonly leverage visual backbones to encode gait silhouettes, achieving successful performance. However, they primarily focus on continuous visual features, overlooking the discrete nature of binary silhouettes that inherently share a discrete encoding space with natural language. Large Language Models (LLMs) have demonstrated exceptional capability in extracting discriminative features from discrete sequences and modeling long-range dependencies, highlighting their potential to capture temporal motion patterns by identifying subtle variations. Motivated by these observations, we explore bridging binary gait silhouettes and natural language within a binary encoding space. However, the encoding spaces of text tokens and binary gait silhouettes remain misaligned, primarily due to differences in token frequency and density. To address this issue, we propose the Contour-Velocity Tokenizer, which encodes binary gait silhouettes while reshaping their distribution to better align with the text token space. We then establish a dual-branch framework termed Silhouette Language Model, which enhances visual silhouettes by integrating discrete linguistic embeddings derived from LLMs. Implemented on mainstream gait backbones, SilLang consistently improves state-of-the-art methods across SUSTech1K, GREW, and Gait3D.

cs.CV↗

arXiv2026

Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models

We identify a fundamental Narrow Policy limitation undermining the performance of autonomous VLA models, where driving Imitation Learning (IL) tends to collapse exploration and limit the potential of subsequent Reinforcement Learning (RL) stages, which often saturate prematurely due to insufficient feedback diversity. Thereby, we propose Curious-VLA, a framework that alleviates the exploit-explore dilemma through a two-stage design. During IL, we introduce a Feasible Trajectory Expansion (FTE) strategy to generate multiple physically valid trajectories and a step-wise normalized trajectory representation to adapt this diverse data. In the RL stage, we present Adaptive Diversity-Aware Sampling (ADAS) that prioritizes high-diversity samples and introduce Spanning Driving Reward (SDR) with a focal style weighting to amplify reward's value span for improving sensitivity to driving quality. On the Navsim benchmark, Curious-VLA achieves SoTA results (PDMS 90.3, EPDMS 85.4) and a Best-of-N PDMS of 94.8, demonstrating its effectiveness in unlocking the exploratory potential of VLA models. Code: https://github.com/Mashiroln/curious_vla.git.

cs.CV↗

arXiv2026

From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving

Vision-Language-Action (VLA) driving augments end-to-end (E2E) planning with language-enabled visual backbones, yet it remains unclear how vision-language models (VLMs) differ from standard vision-only encoders and whether these differences survive policy learning. We study this question under a unified VLM-hidden + diffusion-policy paradigm, comparing multiple VLM families/scales (e.g., InternVL3 and Qwen3VL) with vision-only encoders (e.g., ResNet, ViT, and EVA-CLIP). We ask three questions: how similar are their representations, do residual differences induce meaningful behavioral differences, and how can they improve accuracy-cost trade-offs? We find that VLM and vision-only policies share a substantial common subspace after policy learning, yet both retain non-transferable residual subspaces. Using a Shared-Unique SAE, we show that these residual factors are behaviorally relevant: vision-only encoders are stronger in simple, geometry-dominant scenarios, whereas VLMs are stronger in long-tail, semantically complex, and interaction-heavy cases. The two policy families also exhibit distinct driving styles, with vision-only models being more conservative on average and VLMs more assertive. Exploiting the complementarity between a VLM branch and a ViT branch yields an oracle upper bound of 93.58 PDMS on NAVSIM. We introduce HybridDriveVLA, which runs both branches and uses a learned trajectory scorer for selection, improving PDMS to 92.10 (+1.30 over the VLM baseline), and DualDriveVLA, a fast-slow variant that invokes the VLM in only 15% of scenarios, achieving 91.00 PDMS (+0.20) with about a 1.9x latency speedup over the VLM baseline. Code will be released.

cs.RO↗

arXiv2026

Artificial Entanglement in the Fine-Tuning of Large Language Models

Large language models (LLMs) can be adapted to new tasks using parameter-efficient fine-tuning (PEFT) methods that modify only a small number of trainable parameters, often through low-rank updates. In this work, we adopt a quantum-information-inspired perspective to understand their effectiveness. From this perspective, low-rank parameterizations naturally correspond to low-dimensional Matrix Product States (MPS) representations, which enable entanglement-based characterizations of parameter structure. Thereby, we term and measure "Artificial Entanglement", defined as the entanglement entropy of the parameters in artificial neural networks (in particular the LLMs). We first study the representative low-rank adaptation (LoRA) PEFT method, alongside full fine-tuning (FFT), using LLaMA models at the 1B and 8B scales trained on the Tulu3 and OpenThoughts3 datasets, and uncover: (i) Internal artificial entanglement in the updates of query and value projection matrices in LoRA follows a volume law with a central suppression (termed as the "Entanglement Valley"), which is sensitive to hyper-parameters and is distinct from that in FFT; (ii) External artificial entanglement in attention matrices, corresponding to token-token correlations in representation space, follows an area law with logarithmic corrections and remains robust to LoRA hyper-parameters and training steps. Drawing a parallel to the No-Hair Theorem in black hole physics, we propose that although LoRA and FFT induce distinct internal entanglement signatures, such differences do not manifest in the attention outputs, suggesting a "no-hair" property that results in the effectiveness of low rank updates. We further provide theoretical support based on random matrix theory, and extend our analysis to an MPS Adaptation PEFT method, which exhibits qualitatively similar behaviors.

cs.LG↗

arXiv2025

Chameleon: Benchmarking Detection and Backtracking on Commercial-Grade AI-Generated Videos

The proliferation of AI-Generated Content (AIGC), especially deepfake videos, poses a severe threat to social trust by enabling fraud, privacy violations and disinformation. Existing AI-generated video detection (AGVD) benchmarks focus on open-source model generated videos, yet commercial closed-source models produce more realistic, temporally coherent videos that are underexplored in detection research. To fill this gap, we present Chameleon, a commercial-grade dataset with 1,700 AI-generated videos from 600 real-world sources across three key domains (News, Speech, Recommendation), featuring high resolution, rich annotations and 3D consistency metrics for dynamic scene spatial coherence, shifting detection from face-centric forgery to holistic scene forensics. This benchmark assesses models on two core tasks: accurate AI video detection in real-world conditions and forensic backtracking of original sources. Experimental results reveal critical limitations of existing methods in detecting and backtracking high-fidelity, spatiotemporally consistent videos from commercial closed-source models, highlighting current methods' flawed forensic reasoning and establishing Chameleon as a vital challenge for AIGC security research. The code and data are available at https://github.com/lxixim/Chameleon.

cs.CV↗

arXiv2025

AILuminate: Introducing v1.0 of the AI Risk and Reliability Benchmark from MLCommons

The rapid advancement and deployment of AI systems have created an urgent need for standard safety-evaluation frameworks. This paper introduces AILuminate v1.0, the first comprehensive industry-standard benchmark for assessing AI-product risk and reliability. Its development employed an open process that included participants from multiple fields. The benchmark evaluates an AI system's resistance to prompts designed to elicit dangerous, illegal, or undesirable behavior in 12 hazard categories, including violent crimes, nonviolent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, suicide and self-harm, intellectual property, privacy, defamation, hate, sexual content, and specialized advice (election, financial, health, legal). Our method incorporates a complete assessment standard, extensive prompt datasets, a novel evaluation framework, a grading and reporting system, and the technical as well as organizational infrastructure for long-term support and evolution. In particular, the benchmark employs an understandable five-tier grading scale (Poor to Excellent) and incorporates an innovative entropy-based system-response evaluation. In addition to unveiling the benchmark, this report also identifies limitations of our method and of building safety benchmarks generally, including evaluator uncertainty and the constraints of single-turn interactions. This work represents a crucial step toward establishing global standards for AI risk and reliability evaluation while acknowledging the need for continued development in areas such as multiturn interactions, multimodal understanding, coverage of additional languages, and emerging hazard categories. Our findings provide valuable insights for model developers, system integrators, and policymakers working to promote safer AI deployment.

cs.CY↗

arXiv2025

SearchRAG: Can Search Engines Be Helpful for LLM-based Medical Question Answering?

Large Language Models (LLMs) have shown remarkable capabilities in general domains but often struggle with tasks requiring specialized knowledge. Conventional Retrieval-Augmented Generation (RAG) techniques typically retrieve external information from static knowledge bases, which can be outdated or incomplete, missing fine-grained clinical details essential for accurate medical question answering. In this work, we propose SearchRAG, a novel framework that overcomes these limitations by leveraging real-time search engines. Our method employs synthetic query generation to convert complex medical questions into search-engine-friendly queries and utilizes uncertainty-based knowledge selection to filter and incorporate the most relevant and informative medical knowledge into the LLM's input. Experimental results demonstrate that our method significantly improves response accuracy in medical question answering tasks, particularly for complex questions requiring detailed and up-to-date knowledge.

cs.CL↗

arXiv2024

From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge

Assessment and evaluation have long been critical challenges in artificial intelligence (AI) and natural language processing (NLP). Traditional methods, usually matching-based or small model-based, often fall short in open-ended and dynamic scenarios. Recent advancements in Large Language Models (LLMs) inspire the "LLM-as-a-judge" paradigm, where LLMs are leveraged to perform scoring, ranking, or selection for various machine learning evaluation scenarios. This paper presents a comprehensive survey of LLM-based judgment and assessment, offering an in-depth overview to review this evolving field. We first provide the definition from both input and output perspectives. Then we introduce a systematic taxonomy to explore LLM-as-a-judge along three dimensions: what to judge, how to judge, and how to benchmark. Finally, we also highlight key challenges and promising future directions for this emerging area. More resources on LLM-as-a-judge are on the website: https://llm-as-a-judge.github.io and https://github.com/llm-as-a-judge/Awesome-LLM-as-a-judge.

cs.AI↗

arXiv2024

ClinicalBench: Can LLMs Beat Traditional ML Models in Clinical Prediction?

Large Language Models (LLMs) hold great promise to revolutionize current clinical systems for their superior capacities on medical text processing tasks and medical licensing exams. Meanwhile, traditional ML models such as SVM and XGBoost have still been mainly adopted in clinical prediction tasks. An emerging question is: Can LLMs beat traditional ML models in clinical prediction? Thus, we build a new benchmark ClinicalBench to comprehensively study the clinical predictive modeling capacities of both general-purpose and medical LLMs, and compare them with traditional ML models. ClinicalBench embraces three common clinical prediction tasks, two databases, 14 general-purpose LLMs, 8 medical LLMs, and 11 traditional ML models. Through extensive empirical investigation, we discover that both general-purpose and medical LLMs, even with different model scales, diverse prompting or fine-tuning strategies, still cannot beat traditional ML models in clinical prediction yet, shedding light on their potential deficiency in clinical reasoning and decision-making. We call for caution when practitioners adopt LLMs in clinical applications. ClinicalBench can be utilized to bridge the gap between LLMs' development for healthcare and real-world clinical practice.

cs.CL↗

arXiv2024

Can Knowledge Editing Really Correct Hallucinations?

Large Language Models (LLMs) suffer from hallucinations, referring to the non-factual information in generated content, despite their superior capacities across tasks. Meanwhile, knowledge editing has been developed as a new popular paradigm to correct erroneous factual knowledge encoded in LLMs with the advantage of avoiding retraining from scratch. However, a common issue of existing evaluation datasets for knowledge editing is that they do not ensure that LLMs actually generate hallucinated answers to the evaluation questions before editing. When LLMs are evaluated on such datasets after being edited by different techniques, it is hard to directly adopt the performance to assess the effectiveness of different knowledge editing methods in correcting hallucinations. Thus, the fundamental question remains insufficiently validated: Can knowledge editing really correct hallucinations in LLMs? We proposed HalluEditBench to holistically benchmark knowledge editing methods in correcting real-world hallucinations. First, we rigorously construct a massive hallucination dataset with 9 domains, 26 topics and more than 6,000 hallucinations. Then, we assess the performance of knowledge editing methods in a holistic way on five dimensions including Efficacy, Generalization, Portability, Locality, and Robustness. Through HalluEditBench, we have provided new insights into the potentials and limitations of different knowledge editing methods in correcting hallucinations, which could inspire future improvements and facilitate progress in the field of knowledge editing.

cs.CL↗

arXiv2024

Are LLMs Socially Adaptive? Contrasting Belief Evolution in Large Language Models and Humans

As large language models (LLMs) increasingly engage in complex social interactions, ensuring that their behaviors align with human ethical principles and intentions, known as value alignment, has become a critical scientific challenge. Existing benchmarks often rely on static assessments and fail to capture the longitudinal dynamics of decision-making or the latent cognitive processes driving agent behavior. In this work, we propose FairMindSim, a realistic simulation benchmark rooted in social psychology that evaluates alignment through continuous economic games. To move beyond black-box observations, we introduce the Belief-Reward Alignment Behavior Evolution Model (BREM), a probabilistic framework that formalizes decision-making as a dynamic trade-off between maximizing extrinsic rewards and upholding intrinsic beliefs. We conducted a large-scale comparative study involving 1,017 human participants and ten LLMs, including GPT-5 and Gemini-3-Pro. Our experimental results reveal a capability linked non linear empirical trend in the Third Party Punishment (TPP) game. Mid capability models exhibit rigid and algorithmic aggression that is characterized by over punishment, while frontier models show a convergence of restraint and a shift toward human like leniency as reasoning capabilities scale. Furthermore, using BREM, we decompose agents longitudinal decision dynamics and find that more advanced models better balance conflicting objectives by reducing belief action inconsistency. Our contributions provide a standardized protocol for psychological stress testing and an interpretable mechanism for analyzing the longitudinal evolution of AI alignment in controlled social dilemma settings.

cs.CE↗

arXiv2024

FMBench: Benchmarking Fairness in Multimodal Large Language Models on Medical Tasks

Advancements in Multimodal Large Language Models (MLLMs) have significantly improved medical task performance, such as Visual Question Answering (VQA) and Report Generation (RG). However, the fairness of these models across diverse demographic groups remains underexplored, despite its importance in healthcare. This oversight is partly due to the lack of demographic diversity in existing medical multimodal datasets, which complicates the evaluation of fairness. In response, we propose FMBench, the first benchmark designed to evaluate the fairness of MLLMs performance across diverse demographic attributes. FMBench has the following key features: 1: It includes four demographic attributes: race, ethnicity, language, and gender, across two tasks, VQA and RG, under zero-shot settings. 2: Our VQA task is free-form, enhancing real-world applicability and mitigating the biases associated with predefined choices. 3: We utilize both lexical metrics and LLM-based metrics, aligned with clinical evaluations, to assess models not only for linguistic accuracy but also from a clinical perspective. Furthermore, we introduce a new metric, Fairness-Aware Performance (FAP), to evaluate how fairly MLLMs perform across various demographic attributes. We thoroughly evaluate the performance and fairness of eight state-of-the-art open-source MLLMs, including both general and medical MLLMs, ranging from 7B to 26B parameters on the proposed benchmark. We aim for FMBench to assist the research community in refining model evaluation and driving future advancements in the field. All data and code will be released upon acceptance.

cs.CV↗

arXiv2024

Authorship Attribution in the Era of LLMs: Problems, Methodologies, and Challenges

Accurate attribution of authorship is crucial for maintaining the integrity of digital content, improving forensic investigations, and mitigating the risks of misinformation and plagiarism. Addressing the imperative need for proper authorship attribution is essential to uphold the credibility and accountability of authentic authorship. The rapid advancements of Large Language Models (LLMs) have blurred the lines between human and machine authorship, posing significant challenges for traditional methods. We present a comprehensive literature review that examines the latest research on authorship attribution in the era of LLMs. This survey systematically explores the landscape of this field by categorizing four representative problems: (1) Human-written Text Attribution; (2) LLM-generated Text Detection; (3) LLM-generated Text Attribution; and (4) Human-LLM Co-authored Text Attribution. We also discuss the challenges related to ensuring the generalization and explainability of authorship attribution methods. Generalization requires the ability to generalize across various domains, while explainability emphasizes providing transparent and understandable insights into the decisions made by these models. By evaluating the strengths and limitations of existing methods and benchmarks, we identify key open problems and future research directions in this field. This literature review serves as a roadmap for researchers and practitioners interested in understanding the state of the art in this rapidly evolving field.

cs.CY↗

arXiv2024

Model Attribution in LLM-Generated Disinformation: A Domain Generalization Approach with Supervised Contrastive Learning

Model attribution for LLM-generated disinformation poses a significant challenge in understanding its origins and mitigating its spread. This task is especially challenging because modern large language models (LLMs) produce disinformation with human-like quality. Additionally, the diversity in prompting methods used to generate disinformation complicates accurate source attribution. These methods introduce domain-specific features that can mask the fundamental characteristics of the models. In this paper, we introduce the concept of model attribution as a domain generalization problem, where each prompting method represents a unique domain. We argue that an effective attribution model must be invariant to these domain-specific features. It should also be proficient in identifying the originating models across all scenarios, reflecting real-world detection challenges. To address this, we introduce a novel approach based on Supervised Contrastive Learning. This method is designed to enhance the model's robustness to variations in prompts and focuses on distinguishing between different source LLMs. We evaluate our model through rigorous experiments involving three common prompting methods: ``open-ended'', ``rewriting'', and ``paraphrasing'', and three advanced LLMs: ``llama 2'', ``chatgpt'', and ``vicuna''. Our results demonstrate the effectiveness of our approach in model attribution tasks, achieving state-of-the-art performance across diverse and unseen datasets.

cs.CL↗