Searcharxiv⌕ Search

arXiv subjects

Zhichao Hu

Publications and source records attributed to Zhichao Hu.

At least 19 recordsLinked to original sources

CodeLift: Your Test Suite Is Secretly a Dense SWE Curriculum

The pass-all-test binary reward has become a standard recipe for training software engineering agents, but its sparsity eliminates difference of partially correct patches. The test pass-rate is a natural alternative. However, we find that it is not a reliable indicator of patch quality. A functionality being tested can be covered by either one challenging test or one hundred trivial tests, where the majority might be already satisfied pass-to-pass tests. We introduce CodeLift, a reinforcement learning method that dynamically calibrates pass-rate rewards, turning the test suite into an adaptive curriculum. At each iteration, CodeLift groups tests that kill the same patches and patches that pass the same tests, then gives more weight to test groups that fewer patch groups pass. This reduces the impact of majority simple tests and focuses more on what is still challenging. Within each rollout, CodeLift applies the same scoring rule to repository snapshots, assigning credit to useful actions even when later ones introduce regressions. Experiments on SWE benchmarks show that CodeLift consistently improves pass@1/2/4/8 over binary and pass-rate baselines. Analyses show that CodeLift provides dense, stable learning signals and unlocks the inherent curriculum in the test suite.

cs.LG↗

Benchmarking Coding Agents under Evolving Requirements

Software development with coding agents often begins with a rough request that users refine after trying the implementation. Final correctness alone does not capture the repeated guidance this process requires. We introduce FeatureLoop, a benchmark of 61 human-reviewed tasks from 34 repositories across five primary programming languages and seven software forms. Each task follows one development goal in a persistent workspace through evolution feedback, which adds or clarifies requirements, and completion feedback, which reports failures of stated requirements. Cumulative executable checks track new behavior, repair, and preservation of earlier functionality. We evaluate 12 frontier models and find that the strongest completes 72.1% of tasks with both feedback types. Across models, however, only 15.4% of observed failed stages recover within two corrective revisions. Further analysis reveals incomplete repairs, regressions, and model-specific differences in feedback use and token efficiency, identifying obstacles to reliable, low-intervention development.

cs.SE↗

FeatureArc: Benchmarking Coding Agents on Evolving Features

A software feature usually evolves over time: developers extend its first version as new use cases and requirements arrive, and each extension builds on code written earlier. Most feature benchmarks evaluate a single request on a fixed codebase. Feature evolution requires evaluating successive extensions on an agent's own earlier code and checking whether the feature remains correct. We introduce FeatureArc to evaluate coding agents on evolving features, together with a pipeline that constructs its tasks from repository histories. The benchmark combines dependencies between successive stages of the same feature, explicit requirements and interface contracts, and executable checks across four programming languages. The pipeline validates fail-to-pass (F2P) tests for new behavior and pass-to-pass (P2P) tests for existing behavior on the reference code before and after each stage. We evaluated 10 coding agents on 73 manually reviewed tasks of 2 to 10 stages. Each agent works in one persistent session and workspace without external test feedback or rollback. The evaluation reveals a gap between first-stage and whole-feature success: pooled success falls from 62.7% at the first stage to 28.5% when the whole feature is tested on the final code. Trajectory analysis identifies inherited defects as a common failure mechanism, present in 72.9% of failed later stages. These earlier errors often remain undetected: in stages following a failure, agents leave the earlier error unnoticed about 95% of the time, and their self-tests rarely revisit earlier stages. Evolving a feature autonomously, without human checks between stages, therefore remains a substantial challenge for current coding agents. At Tencent, the pipeline extends internal evaluations to this setting and guides the development of Hunyuan, Tencent's family of large language models.

cs.SE↗

AgentPerturb: Evaluating LLM Agent Robustness under Controlled Interaction Perturbations

Real-world tasks are often complex and noisy, posing substantial challenges for LLM agents. Yet existing evaluations lack a systematic framework for jointly characterizing potential perturbation targets and types. We use this two-dimensional view to conduct a unified study that quantifies the robustness of current frontier models, traces the mechanisms behind perturbation-induced failures, and draws implications for robustness-oriented training. Through controlled evaluations on APEX-Agents and Toolathlon, we find that perturbations consistently degrade performance, while the dominant vulnerabilities vary across task environments. Trajectory-level analysis reveals that agents rarely recognize altered inputs and often propagate corrupted information without verification. A focused cost analysis further shows that successful perturbed trajectories can require substantially more computation. These findings motivate training objectives that cover diverse forms of noise and reward verification, uncertainty awareness, and recovery. We provide a rigorous benchmark-construction pipeline. Code and data will be released at https://github.com/xuyj233/AgentPerturb.

cs.AI↗

DRILL: Training World Models to Improve Policies, Not Predict Pixels

World models are increasingly used to reduce environment interaction when post-training vision-language-action (VLA) policies with reinforcement learning. Yet they are typically trained to predict pixels one step ahead, even though their role here is not merely to render visually faithful videos but to produce learning signals from imagined rollouts that improve the policy. We show that this mismatch is substantial: lower pixel MSE can produce worse policy gradients, and a world model trained for policy utility can yield better downstream policies despite 17% higher pixel MSE on the next frame. We propose the Downstream-Return Imagination Learning Loop (DRILL), a bilevel framework that updates the world model to maximize the policy's return after an inner GRPO update on imagined rollouts. DRILL has two instantiations. DRILL-VWI is a closed-form surrogate, requiring no additional simulator rollouts, that reweights prediction errors by policy visitation, the group-relative GRPO advantage, and model confidence. DRILL-IMG is a full meta-gradient method that differentiates through the inner update using truncated Hessian-vector products. We show that DRILL-VWI recovers the local one-step component of the DRILL-IMG meta-gradient under score-compatibility and small-residual assumptions. Across five manipulation simulators and two VLA backbones, DRILL-IMG raises mean success rate by 13.1 points over WMPO and 7.9 points over RLVR-World, and complements RLVR pretraining rather than competing with it. More broadly, we believe video world models for VLA reinforcement learning can be productively trained by the policy gradients they produce, not only by pixel-level prediction.

cs.CV↗

UserSWE: Evaluating Coding Agents in Multi-Turn Conversations with Dynamic Users

As coding agents become more capable, developers increasingly use them to fix bugs through multi-turn interactions. Existing benchmarks do not adequately cover this scenario. Most present a complete issue report in a single turn without interaction. Their evaluation can produce both false positives and false negatives, accepting incorrect patches or rejecting correct ones. We introduce UserSWE for evaluating coding agents' ability to fix repository-level bugs through multi-turn interactions with dynamic users and execution-grounded verdicts at every turn. Its construction pipeline turns repository history and existing single-turn benchmarks into multi-turn tasks. At each turn, information pools constrain what the simulated user can reveal, while a persona chain sampled from a transition model calibrated on 20,859 development sessions governs how it speaks. Tests run after every turn, and a pair of agents from different model families writes and gates each verdict, using test results as primary evidence and model judgment only for aspects the tests cannot resolve. We evaluate eleven agents on 72 tasks from 42 repositories and find that agent rankings change with the interaction budget. Claude Fable 5.1 resolves the most tasks in one turn, GPT-6-Astra resolves the most by the eighth turn, and Claude Opus 5 leads at most intermediate budgets. These results indicate that single-turn performance is insufficient to characterize interactive coding ability.

cs.AI↗

VisualTrust: Video Reasoning under On-Screen Claim Framing

On-screen text can describe what a video shows, or it can frame what the video is supposed to mean. Current benchmarks treat both alike, testing whether a model can read the text rather than whether it should trust it. We study claim deference, multimodal large language models' tendency to let an on-screen claim override conflicting visual evidence. We introduce VisualTrust, a counterfactual benchmark with 5,063 cases across 14 claim-fact classes. Each case pairs a question with three versions of the same video: one with no claim, one with an accurate claim, and one with a misleading claim that alters a single grounded fact. The option the misleading claim supports is the trap answer. Across seven open-weight and four proprietary models, mean accuracy falls from 90.1% without a claim to 51.9% with a misleading claim. Among initially correct responses, 42.8% flip to the trap answer and only 0.6% become other errors. Strong no-claim accuracy does not ensure resistance. We also introduce DEAR (Decoupled Evidence Arbitration and Reading), a training-free agent for video question answering that separates evidence arbitration from reading. An independent check reports conflicting evidence, a selector weighs it without a fixed modality priority and blinds what it withholds, and the reader answers from the resulting video in a fresh context. On 500 videos across three proprietary models, DEAR raises accuracy under misleading claims from 47.2--64.0% to 81.8--85.2% and reduces trap-answer selection from 34.4--50.8% to 11.6--15.0%, while preserving no-claim accuracy.

cs.CV↗

Q-Save: Towards Scoring and Attribution for Generated Video Evaluation

Evaluating AI-generated video (AIGV) quality hinges on three crucial dimensions: visual quality, dynamic quality, and text-video alignment. While numerous evaluation datasets and algorithms have been proposed, existing approaches are constrained by two limitations: the absence of systematic definitions for evaluation dimensions, and the isolated treatment of the three dimensions in separate models. Therefore, we introduce Q-Save, a holistic benchmark dataset and unified evaluation model for AIGV quality assessment. The Q-Save dataset contains nearly 10,000 video samples, each annotated with Mean Opinion Scores (MOS) and fine-grained attribution explanations across the three core dimensions. Leveraging this attribution-annotated dataset, we train the proposed Q-Save model, which adopts the SlowFast framework to balance accuracy and efficiency, and employs a three-stage training strategy with Chain-of-Thought (COT) formatted data: Supervised Fine-Tuning (SFT), Grouped Relative Policy Optimization (GRPO), and a final SFT round for stability, to jointly perform quality scoring and attribution generation. Experimental results demonstrate that Q-Save achieves superior performance in AIGV quality prediction while providing interpretable justifications. Code and dataset will be released upon publication.

cs.CV↗

Modality Fault Lines: Structural Corruptions Reveal Fragile Omni-Modal Reasoning

Omni-modal large language models are increasingly evaluated on clean text--vision--audio inputs, where every channel is present, synchronized, and readily interpretable. Such scores are often taken as evidence of robust cross-modal fusion, but clean evaluation cannot tell whether success depends on stable cross-modal structure or on cues sufficient only in intact inputs. To address this gap, we define a modality fault line: a boundary at which model behavior becomes unstable when a modality remains present and human-interpretable, but its internal evidence structure is perturbed. We introduce SCEval (Structure-Corruption Evaluation) a diagnostic evaluation protocol that keeps the question, answer space, and modality channels fixed while applying controlled structural corruptions to text, vision, and audio individually and jointly. Built from $273$ human-verified tri-modal examples from Social-IQ, OmniBench, and VALOR, SCEval evaluates $15$ proprietary and open-source omni-modal systems. The results show that structural corruption lowers clean accuracy, text--vision damage forms the most stable shared fault line, and multi-modal degradation is non-additive rather than a simple function of the number of corrupted modalities. Clean omni-modal accuracy therefore does not establish that a model will remain reliable when cross-modal evidence becomes structurally unreliable.

cs.CL↗

JPO: Juris Policy Optimization for Structured Legal Reasoning in Criminal Judgment Prediction

Criminal judgment prediction requires models to infer statutory articles, charges, and sentencing outcomes from case facts. Unlike standard classification tasks, it involves a structured reasoning process in which statutes should be matched with facts, charges should be justified by statutes, and sentencing outcomes should remain consistent with charges. Existing approaches optimize final labels, and while some have attempted to evaluate reasoning quality, their evaluations are indirect, often relying on LLM-generated rubrics that reflect model-internal preferences rather than the inherent logical structure of legal adjudication. We propose Juris Policy Optimization (JPO), a post-training framework for structured legal reasoning in Chinese criminal judgment prediction. JPO first uses teacher-generated rationales to supervise a standardized four-step reasoning process, and then applies reinforcement learning with a composite reward over legal prediction quality, reasoning structure completeness, and cross-step consistency. JPO further introduces token-level advantage reweighting and adaptive clipping for legally salient reasoning segments. Experiments on multiple open-source language models and three Chinese legal benchmarks show that JPO consistently improves both judgment prediction and reasoning quality over supervised fine-tuning and reinforcement learning baselines.

cs.CL↗

DramaChain Bench: An End-to-End Benchmark for Short-Drama Generation

Commercial short-drama production follows a multi-stage chain: script, storyboard, keyframe imagery, shot-level video, and the finished short drama. Most existing benchmarks evaluate solely the video-generation stage using pre-authored inputs instead of real upstream pipeline outputs. This leaves two critical questions unanswerable: whether each stage adheres to the original script intent (rather than only its immediate input prompt), and whether disparate shots remain coherent after assembly into multi-episode releases. We present DramaChain Bench, the first short-drama benchmark that evaluates every stage of the complete production chain. It is built upon three in-house systems sharing one dimension system, DramaChain Dimensions: five evaluation axes instantiated at every stage, resolving into 63 leaf dimensions. DramaChain Agent is calibrated against commercial short-drama platforms in both workflow and finished short-drama quality, enabling stage-wise fair comparison across models. DramaChain Labeling System has each of the 5,785 items scored independently by three professional annotators, with all defects spatio-temporally localised and selected from a predefined defect list. This process produces 17,488 valid scores and 255,925 traceable attribution records. The human annotations confirm that upstream defects cascade across the pipeline, demonstrating that final episode quality is not governed by video generation alone. DramaChain Agentic Judge then scores every leaf dimension automatically, gathering evidence over multiple agentic rounds before judging against a per-item checklist; it reproduces the model ranking at a mean PLCC of 0.918, enough to admit new models at no annotation cost.

cs.AI↗

VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes

Frontier multimodal large language models (MLLMs) have been reported to achieve over 90\% accuracy on fine-grained perception benchmarks. However, such scores do not necessarily imply faithful use of visual evidence. Prior studies have identified three shortcuts that inflate benchmark performance. First, linguistic priors and lexical cues in questions often enable models to infer plausible answers without seeing the image. Second, coarse global semantics from the visual encoder can bypass fine-grained local details. Third, in some ``think-with-images'' benchmarks, corrupting the intermediate images returned by visual tools barely affects the final answer. These findings suggest that higher input resolution or larger question pools alone do not elicit genuine active visual search. To address this, we introduce VisualNeedle, a challenging, information-dense, and fine-grained benchmark for scenes where critical evidence is spatially constrained to minute regions and not discernible at a glance. We further propose a counterfactual crop-black setting, which replaces crops returned by tools with black images of the same size, to test whether tool-enabled performance truly relies on intermediate visual evidence.We evaluate 9 prominent MLLMs across four settings: text-only, without tools, with tools, and crop-black. Text-only accuracy stays below 10\%, while accuracy without tools remains below 20\%. The best tool-enabled model reaches only 56.00\%, still trailing the 63.00\% human majority-vote accuracy. These results reveal persistent limitations in fine-grained visual search, while the crop-black ablation confirms that success on VisualNeedle hinges on genuine intermediate visual evidence.

cs.CV↗

Can LLM-as-a-Judge Reliably Verify Rubrics in Agentic Scenarios?

Rubric-based scoring has become a widely used paradigm in model evaluation, typically with LLM-as-a-Judge (LaaJ) for rubric scoring. However, the reliability of LaaJ for rubric scoring remains underexplored. This concern is especially pronounced in agentic scenarios, where long, complex outputs further challenge reliable scoring. To address this, we conduct a systematic meta-evaluation of LaaJ reliability for rubric verification. We introduce RuVerBench, the first benchmark for assessing LaaJ reliability in rubric verification for agentic scenarios. RuVerBench covers two prevalent agentic domains, deep research and agentic coding, with 2,458 instances, each containing a model-generated output, a rubric, and a human-annotated label indicating whether the output satisfies the rubric. Using RuVerBench, we evaluate numerous frontier LLMs and find that even the most advanced models achieve strong performance but still exhibit substantial noise. We further analyze the impact of key LaaJ strategies, including prompt design, batching, and majority voting, on rubric verification. We find that weaker models are more sensitive to prompt variations, batched verification presents a trade-off between accuracy and efficiency, and majority voting yields effective but diminishing returns. We have released our dataset and code to facilitate future research: https://github.com/THU-KEG/RuVerBench.

cs.CL↗

RepoProbe: Benchmarking Architecture-Aware Repository Comprehension with Checklists

The integration of Large Language Models (LLMs) into software engineering has shifted the focus from function-level generation to repository-scale assistance. However, existing benchmarks largely rely on bug reports from GitHub Issues, which often allow models to bypass genuine understanding via pattern matching on error logs. This misalignment under-measures Edit Bias, which refers to premature generation, where models prematurely propose code modifications instead of understanding the existing repository architecture. Furthermore, current LLM-as-a-Judge scalar scoring suffers from high variance and low interpretability. This work introduces RepoProbe, a novel benchmark for evaluating repository-level code understanding through open-ended Q&A using GitHub Discussions, which focuses on open-ended architectural inquiries rather than defect reporting. To ensure rigorous evaluation, we propose a Checklist-Based Verification Protocol that decomposes answers into atomic, verifiable facts, thereby replacing subjective ratings with objective verification. Our evaluation of state-of-the-art (SOTA) LLMs reveals a persistent gap between high clarity and evidencegrounded technical correctness. It also quantitatively confirms the prevalence of edit bias, in which models prioritize code generation instead of architectural analysis. Finally, we demonstrate that our verification protocol significantly improves evaluation reliability compared to traditional evaluations with scalar scoring.

cs.SE↗

TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories

LLM-based agentic systems have shown remarkable capabilities in complex domains, while suffering from cascading errors and difficulty in debugging. Critical error detection aims to locate the earliest error step in a failed trajectory that is responsible for the final failure. However, progress faces two main challenges. First, long trajectories make it difficult to identify individual errors, since the evidence for judging a step may be scattered across distant instructions, observations, and prior context. Second, failed trajectories often contain multiple local errors with different downstream effects, only some of which remain responsible for the final failure. In this work, we propose TrajDebug, an error-lifecycle tracing framework that addresses long-trajectory error discovery with multi-granularity history compression and evidence-based error identification, and supports critical attribution by tracing each error's resolution status and terminal impact. We further construct TrajErrBench, a benchmark of 486 manually annotated failed trajectories from Tau2Bench and SWE-Bench Pro, covering realistic tool-use and coding scenarios. Experiments across diverse agent benchmarks show that TrajDebug achieves the best overall performance over existing baselines, and application studies further demonstrate that its diagnoses provide actionable feedback for improving downstream agent success. We will release the codes and data to facilitate further research.

cs.AI↗

Skill-Use: Can LLMs Actually Use Skills in Agentic Harnesses?

Large language model (LLM) agents increasingly rely on skills, structured documents that specify when to act, which procedure to follow, and which tools are allowed. Existing evaluations mostly judge the quality of a skill or its contribution to task success, leaving unexamined whether an agent can recognize a relevant skill and apply it on its own. We introduce Skill-Use, a benchmark that evaluates skill use under progressive disclosure, where an agent sees only a skill's name and short description and must retrieve the full procedure before following it. Skill-Use separates three facets of skill use. Trigger measures whether the agent invokes the relevant skill, Compliance measures how faithfully it follows the prescribed procedure, and Boundary measures whether it avoids forbidden operations. A Skill-Use (SU) score combines the three and credits execution only after the skill is triggered. Skill-Use pairs 79 real skills with 177 executable tasks across nine domains, each grounded in real files, run in an isolated Docker sandbox, and scored by a trajectory-based rubric. Evaluating eight LLMs under two agent harnesses, we find that reliable skill use remains out of reach, as the strongest configuration reaches an SU of only 0.613. Triggering and procedural compliance fail as independent bottlenecks, and both scores and model rankings shift with the harness, so skill use behaves as a capability conditioned on the harness rather than a fixed property of the model.

cs.CL↗

When Search Agents Should Ask: DiscoBench for Clarification-Aware Deep Search

Search agents powered by large language models (LLMs) are increasingly used to solve complex information-seeking tasks, requiring multi-step retrieval and reasoning to fulfill user goals. However, existing benchmarks often assume that user queries are complete and explicit, overlooking the fact that real-world search requests are frequently vague, underspecified, or even factually incorrect. In deep search scenarios, such ambiguity can propagate along multi-step reasoning chains and lead agents toward incorrect search trajectories. To address this gap, we introduce DiscoBench, a benchmark for clarification-aware deep search, designed to evaluate whether search agents can proactively identify ambiguity, ask effective clarification questions, and recover correct reasoning paths through user interaction. DiscoBench contains 211 samples and 463 ambiguity instances across 11 real-world domains, covering four ambiguity types. We further design a user simulator for multi-turn interaction and evaluate model performance from four perspectives: task utility, ambiguity detection, interaction strategy, and cost efficiency. Experiments on representative LLMs show that ambiguity detection and effective clarification are distinct capabilities, and that repeatedly searching instead of asking for clarification often performs worse than direct guessing, highlighting a critical gap between retrieval ability and interactive problem-solving in current search agents.

cs.CL↗

TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment

Recent studies have demonstrated the efficacy of integrating Group Relative Policy Optimization (GRPO) into flow matching models, particularly for text-to-image and text-to-video generation. However, we find that directly applying these techniques to image-to-video (I2V) models often fails to yield consistent reward improvements. To address this limitation, we present TAGRPO, a robust post-training framework for I2V models inspired by contrastive learning. Our approach is grounded in the observation that rollout videos generated from identical initial noise provide superior guidance for optimization. Leveraging this insight, we propose a novel GRPO loss applied to intermediate latents, encouraging direct alignment with high-reward trajectories while maximizing distance from low-reward counterparts. Furthermore, we introduce a memory bank for rollout videos to enhance diversity and reduce computational overhead. Despite its simplicity, TAGRPO achieves significant improvements over DanceGRPO in I2V generation. The deliverables will be updated at https://tagrpo.github.io/ .

cs.CV↗