arXiv subjects

Xuezhi Cao

Publications and source records attributed to Xuezhi Cao.

At least 19 recordsLinked to original sources

arXiv2026

DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement

Online agent deployments produce abundant execution traces, while task-specific verification and expert annotation are costly to scale. We study how to distill these traces into reusable feedback without post-hoc outcome labels, drawing on their evidence of local progress, recovery, and unfinished requirements. We introduce DENSE (Distilling Evidence from Nested Subtask Executions), which organizes this evidence into evidence-grounded nested shortcut trees. DENSE compresses redundant attempts, reconciles issues across levels using recovery evidence, and summarizes completed branches while expanding unresolved ones, linking reusable progress to remaining obligations. We introduce REFIT, a source-paired protocol comparing feedback from shared initial trajectories under post-hoc outcome blindness, with environments and model contexts reset for fresh attempts at the same tasks. On Terminal-Bench 2.1, DENSE achieves the highest strict pass rate among tested non-privileged feedback methods across four recipient models. Relative to initial executions, strict pass rate improves by 7.12-15.64 pp, with 19.0-43.6% fewer observed recipient tokens in reruns. GPT-5.5 ablations support combining nested subtask analysis with shortcut construction and issue reconciliation. These findings point toward agent self-refinement through evidence-grounded trajectory reuse with less reliance on external supervision.

cs.AI↗

arXiv2026

2AM: Grounding Agent-Side Memory as Guidance for Steerable Action Models in Long-Horizon Manipulation

Long-horizon robot manipulation requires memory, but not necessarily inside the action policy. To address such tasks, current agentic systems often combine VLAs with planners and geometric tools, sometimes using additional depth or calibrated geometry. These systems confound attribution: gains may come from richer observations or alternative motor tools, while failures may stem from either the policy or an under-specified language interface. We isolate this question through a deliberately constrained design: less tool breadth, but greater interface bandwidth. 2AM makes a multimodal Agent the sole holder of task memory and a single RGB-based, episodically stateless Action Model the sole executor of task-relevant motion. The Agent compiles interaction history into subtask language and optional 2D grasp, place, and move hints that bind its physical intention at different time scales. To teach this steerability to the VLA, we augment demonstrations with structured hint labels and train under condition dropout, spatial noise, and temporal jitter to tolerate imperfect Agent outputs. On LIBERO-Mem, without depth, online geometry, or planner-based object motion, 2AM reaches 76.3% average completion, a 61.5-point improvement over the strongest reported baseline of 14.8%, together with 63.0% relaxed and 11.8% strict success. These results show that task memory can remain Agent-side. They further show that Action Model capability depends not only on what the policy has learned, but on how precisely the Agent can steer it.

cs.RO↗

arXiv2026

AdaR: A Framework for Equipping LLMs with Adaptive Reasoning

Mathematical reasoning is a primary indicator of large language models (LLMs) intelligence. However, existing LLMs exhibit failures in robustness and generalization. This paper attributes these deficiencies to spurious reasoning, wherein generated reasoning traces are driven by superficial correlations, leading models to blindly reproduce memorized patterns from the training data. To address this challenge, we propose the AdaR framework to equip LLMs with adaptive reasoning, wherein models establish correct correlations between query templates and reasoning processes. AdaR automatically synthesizes logically equivalent queries by varying variable values and trains models using Reinforcement Learning with Verifiable Rewards (RLVR) to penalize spurious logic while encouraging adaptive logic. To ensure data quality, we extract the problem-solving logic from the original query, generate the corresponding answer through code execution, and then apply a sanity check. Experimental results demonstrate that AdaR achieves substantial improvements in mathematical reasoning while maintaining high data efficiency. Furthermore, even advanced LLMs still exhibit deficiencies in robustness and generalization, which our work effectively mitigates. Our project is available at https://github.com/NJUNLP/AdaR.

cs.AI↗

arXiv2026

Contrastive Reinforced Policy Optimization via Privileged Self-Distillation

Recent advances in post-training Large Language Models (LLMs) increasingly rely on Reinforcement Learning with Verifiable Rewards (RLVR) or On-Policy Self-Distillation (OPSD). While OPSD provides dense, logit-level supervision, it inherently suffers from exposure bias due to the privileged information of the self-teacher. In multi-turn agentic settings, this leads to reasoning route convergence and the loss of clear optimization directions. To tackle these challenges, we introduce Contrastive Reinforced Policy Optimization (CRPO), which reformulates agentic OPSD from a contrastive learning perspective. By leveraging predictive entropy to distinguish between positive positions (reflective exploration) and negative positions (exposure bias), CRPO conducts group-wise contrast to preserve reliable, fine-grained optimization signals. Extensive evaluations across 13 challenging reasoning and deep-search benchmarks demonstrate that CRPO consistently outperforms existing reinforcement learning and self-distillation baselines, significantly enhancing training stability and generalization in long-horizon interactions.

cs.LG↗

arXiv2026

ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents

As LLM-based agents are deployed in complex, multi-step workflows, a critical evaluation gap has emerged: most existing benchmarks judge only final outcomes, unable to distinguish reliable reasoning from lucky success or attribute failures to specific process deficiencies, hindering attribution in long-horizon tasks. In this work, we present ClawTrack, a dual-assessment benchmark that simultaneously measures what an agent achieves (Task Score) and how it achieves it (Process Score). ClawTrack comprises 320 tasks across 8 domains with 25+ deterministic mock services. A Process Grader scores each reasoning turn along four dimensions (goal alignment, efficiency, information utilization, and result verification), anchored by 12,541 task-specific rubric items. Evaluating 21 models over 16,000+ trials, we find that: (1) process scores effectively attribute success and failure to specific reasoning dimensions, filtering lucky passes invisible to outcome-only evaluation; (2) the four dimensions are complementary, with result verification as the systematic bottleneck; (3) the framework is robust to evaluator choice across different judge LLMs; and (4) process-based trajectory filtering yields consistent post-training improvements across model scales.

cs.LG↗

arXiv2026

ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders

The recent emergence of vibe-coding workflows is changing what coding agents are expected to do. Instead of merely completing code under fully specified instructions, agents are increasingly expected to transform incomplete product intent into working software by combining various abilities including planning, requirement clarification, tool use, debugging, and repository-level construction. Yet existing benchmarks have not fully caught up with this shift, evaluating agents on static, fully specified tasks. In this paper, we introduce ICAE-Bench, a benchmark for evaluating coding agents under interactive project-building settings. The basic idea is to start from a fuzzy product requirement, simulating the dynamic paradigm with an automated User Agent. To make this setting both realistic and evaluable, ICAE-Bench introduces three key designs. First, to avoid the ambiguity of unconstrained fuzzy requirements, each task derives ambiguity from a precise real open-source repository with executable behavior. Second, to ensure high-quality and reproducible user simulation, ICAE-Bench grounds interaction through User Agent Data, allowing the User Agent to reveal hidden constraints without inventing new requirements or leaking implementation artifacts. Third, to evaluate open-ended repositories fairly, ICAE-Bench uses standardized black-box tests together with multi-dimensional diagnostics, including functional correctness, semantic and API similarity, structural fidelity, design quality, and interaction quality.

cs.AI↗

arXiv2026

OPERA: Aligning Open-Ended Reasoning via Objective Perplexity-based Reinforcement Learning

Reinforcement Learning (RL) has enabled LLMs to excel in objective reasoning tasks such as mathematics and code generation. However, applying RL to open-ended tasks, such as creative writing, remains challenging because LLM-as-a-judge reward models often exhibit stylistic biases and positional inconsistencies, leading to unstable supervision. To address this, we propose OPERA (Objective Perplexity-based Reflective Alignment), which replaces unreliable external judges with intrinsic rewards derived from perplexity dynamics. Specifically, we derive an intrinsic reward signal from perplexity dynamics, quantifying uncertainty reduction at critical reflective states. During the cold-start phase, we introduce a data synthesis method that leverages carefully designed guiding words to generate diverse reasoning traces, along with perplexity-prioritized rollouts that utilize internal log-probabilities to identify logically consistent reasoning branches. This pipeline yields a large-scale dataset comprising 20,000 high-quality reasoning trajectories. Empirical evaluations consistently demonstrate the scalability and efficacy of our approach in alignment for open-ended tasks. Implementing OPERA on Qwen3-8B establishes a new state-of-the-art among open-source models, achieving parity with or surpassing proprietary models like Gemini2.5 and MiniMax-M2.5 in some open-ended tasks. The code is available at https://github.com/pangpang-xuan/OPERA.

cs.CL↗

arXiv2026

DailyReport: An Open-ended Benchmark for Evaluating Search Agents on Daily Search Tasks

Search Agents (SAs) typically leverage large language models (LLMs) to support complex information-seeking tasks by autonomously exploring web sources and synthesizing information into comprehensive responses. For SAs evaluation, prior benchmarks mainly focus on specialized tasks that are unlikely to arise in real-world user scenarios. Moreover, their reliance on coarse task-level rubrics often limits evaluation interpretability. To bridge this gap, we introduce DailyReport, an open-ended benchmark to evaluate SA capabilities on daily search tasks. It contains 150 open-ended tasks with 3,546 associated rubrics, capturing widely discussed and timely information demands of real-world users. Each task is decomposed into subtasks and evaluated with cascade rubrics across disentangled dimensions. Through cascade performance attribution and user-centric aggregation, we derive highly interpretable scores for each dimension, along with a user preference score. Our results on 17 agentic systems show that current systems still fall short of users' expectations. To facilitate future research, our dataset and code are made publicly available at https://github.com/AGI-Eval-Official/DailyReport.

cs.AI↗

arXiv2026

Asuka-Bench: Benchmarking Code Agents on Underspecified User Intent and Multi-Round Refinement

Existing code-generation benchmarks score a single mapping from a complete prompt to a one-shot output. However, real web development is different. Users seldom write a full spec at the start; many requirements only become clear once they look at an intermediate result and react to it. We present Asuka-Bench, a benchmark that pairs underspecified user intent with multi-round refinement, grounded in browser-rendered behavior. Each task is resolved through a closed loop: a Code Agent generates a web project, a UI Agent executes test cases on the deployed site, and a User LLM turns evaluation outcomes into natural-language feedback for the next round. The benchmark comprises 50 web tasks with 784 evaluation criteria and 2402 expected outcomes. We benchmark 8 LLMs across 2 agent frameworks. The results separate models clearly: weighted Task Pass Rate varies by 38 percentage points and models also differ substantially in their ability to repair from feedback. Asuka-Bench is also far from saturated: even the strongest model completes only 52% of projects after three rounds.

cs.SE↗

arXiv2026

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems

Self-improving language agents are typically evaluated in isolation: an agent attempts a task, receives feedback, and iteratively refines its own behavior. Yet agents increasingly operate alongside peers whose strategies and outcomes are publicly visible. This raises an under-studied question: when does shared experience produce improvements that self-improvement alone cannot achieve? We introduce SAGE (Social Agent Group Evolution),an evaluation framework that compares two compute-matched conditions: SocialEvo, where agents from five distinct model families co-evolve with access to all peers' histories; and SelfEvo, where each agent receives the same number of task attempts but sees only its own past, which is conventional in self-improving agent studies. We instantiate SAGE in three arenas: open-ended ML research, long-horizon economic planning, and strategic multiplayer play, evaluated across multiple evolutionary rounds. We find that group history is not a universal amplifier: the strongest agent does not exceed its self-evolution ceiling. However, agents that plateau under self-improvement can achieve significant breakthroughs when peer experience is available. In competitive settings, counterfactual controls reveal that agents improve generally rather than developing opponent-specific strategies. Across different forms of shared history, filtered peer traces and reflective summaries often outperform raw logs, indicating that social gains depend on abstraction rather than exposure volume. These findings reveal that peer-history gains are agent-specific, arena-dependent, and contingent on the capacity to abstract transferable knowledge from public traces.

cs.AI↗

arXiv2026

ATLAS: All-round Testing of Long-context Abilities across Scales

Long-context language models now advertise context windows up to millions of tokens, yet evaluations typically report a single length or a narrow task family, masking two failure modes: performance can collapse as length grows, and strong retrieval need not transfer to downstream use. We present ATLAS, a benchmarking framework that redefines long-context evaluation as length-dependent capability profiling. ATLAS contributes three methodological principles:(i) a layered taxonomy separating foundational operations from application workloads so failures can be attributed, (ii) length-aware AUC scoring that integrates score-length curves over a fixed 8K-1M grid, replacing single-point metrics with full degradation profiles, and (iii) ATLAScore, a harmonic-mean aggregate over taxonomy categories that penalizes imbalanced profiles, with end-to-end uncertainty propagation from subset scores through the nonlinear final aggregate. We instantiate the framework across eight capability dimensions with nine auditable components and 6,438 instances, and evaluate 26 models. Gemini-3.1-Pro-Preview leads at 128K, Claude-Opus-4.6 leads at 1M. Rankings reshuffle substantially between ATLASscore@8K-128K and ATLASscore@8K-1M: 7 models move by at least two ranks, and the two taxonomy layers share only 61% of cross-model variance, with individual rank gaps up to 12 positions. These results support reporting long-context quality by capability and length, not by a single headline score.

cs.CL↗

arXiv2026

WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation

Interactive world models are advancing rapidly, yet existing benchmarks cover only part of the required competencies, leaving no unified standard for systematic evaluation. To fill this gap, we introduce WBench, a comprehensive multi-turn benchmark for interactive world model evaluation along five dimensions, namely video quality, setting adherence, interaction adherence, consistency, and physics compliance. WBench contains 289 test cases and 1,058 interaction turns, where each case specifies a world setting and a multi-turn interaction sequence, covering diverse scenes, styles, subjects, and both first- and third-person perspectives, together with four interaction types, including navigation, subject action, event editing, and perspective switching. For navigation, WBench unifies text, 6-DoF pose, and discrete-action control, enabling evaluation of models with different native input interfaces. Evaluation uses 22 automatic sub-metrics that combine specialist vision models with large multimodal models, and all metrics are validated against human judgments. Across 20 state-of-the-art models, we find that no single model performs strongly across all dimensions. We provide detailed diagnostic insights into the characteristic strengths, weaknesses, and open challenges of each model. Code and data are available at https://github.com/meituan-longcat/WBench.

cs.CV↗

arXiv2026

AgentEscapeBench: Evaluating Out-of-Domain Tool-Grounded Reasoning in LLM Agents

As LLM-based agents increasingly rely on external tools, it is important to evaluate their ability to sustain tool-grounded reasoning beyond familiar workflows and short-range interactions. We introduce AgentEscapeBench, an escape-room-style benchmark that tests whether agents can infer, execute, and revise novel tool-use procedures under explicit long-range dependency constraints. Each task defines a directed acyclic dependency graph over tools and items, requiring agents to invoke real external functions, track hidden state revealed incrementally, propagate intermediate results, and submit a deterministically verifiable final answer. AgentEscapeBench includes 270 instances across five difficulty tiers and supports fully automated evaluation. Experiments with sixteen LLM agents and human participants show that performance drops sharply as dependency depth increases: humans decline from 98.3% success at difficulty-5 to 80.0% at difficulty-25, while the best model drops from 90.0% to 60.0%. Trajectory analysis attributes model failures mainly to breakdowns in long-range state tracking, clue adherence, and intermediate-result propagation. These findings suggest that current agents can often handle local tool use but still struggle with deep contextual dependencies. We hope AgentEscapeBench can serve as a diagnostic testbed for measuring current agent capabilities and informing future training efforts toward more robust general-purpose reasoning, action, and adaptation.

cs.AI↗

arXiv2026

SWE-Cycle: Benchmarking Code Agents across the Complete Issue Resolution Cycle

As autonomous code agents move toward end-to-end software development, evaluating their practical autonomy becomes critical. Current benchmarks hide friction by testing agents in pre-configured environments, and their static evaluation pipelines frequently fail when parsing fully autonomous trajectories. We address these limitations with SWE-Cycle, a benchmark of 489 rigorously filtered instances. SWE-Cycle evaluates agents across three isolated tasks, including environment reconstruction, code implementation, and verification test generation, as well as an end-to-end FullCycle task that integrates all three. The FullCycle task requires agents to work autonomously in a bare repository without human scaffolding. To reliably assess these complex execution paths, we developed SWE-Judge. By combining static code review with dynamic testing, this execution-capable evaluation agent accurately verifies functional correctness and eliminates the systematic measurement errors of traditional static parsers. We evaluate code agents powered by six state-of-the-art LLMs across these four tasks. The results reveal a sharp drop in solve rates when transitioning from isolated tasks to FullCycle execution, exposing critical bottlenecks in handling cross-phase dependencies and maintaining code quality. Together, SWE-Cycle and SWE-Judge provide a comprehensive framework for accurately measuring the end-to-end capabilities of autonomous software agents.

cs.SE↗

arXiv2026

LARY: A Latent Action Representation Yielding Benchmark for Generalizable Vision-to-Action Alignment

While the shortage of explicit action data limits Vision-Language-Action (VLA) models, human action videos offer a scalable yet unlabeled data source. A critical challenge in utilizing large-scale human video datasets lies in transforming visual signals into ontology-independent representations, known as latent actions. However, the capacity of latent action representation to derive robust control from visual observations has yet to be rigorously evaluated. We introduce the Latent Action Representation Yielding (LARY) Benchmark, a unified framework for evaluating latent action representations on both high-level semantic actions (what to do) and low-level robotic control (how to do). The comprehensively curated dataset encompasses over one million videos (1,000 hours) spanning 151 action categories, alongside 620K image pairs and 595K motion trajectories across diverse embodiments and environments. Our experiments reveal two crucial insights: (i) General visual foundation models, trained without any action supervision, consistently outperform specialized embodied latent action models. (ii) Latent-based visual space is fundamentally better aligned to physical action space than pixel-based space. These results suggest that general visual representations inherently encode action-relevant knowledge for physical control, and that semantic-level abstraction serves as a fundamentally more effective pathway from vision to action than pixel-level reconstruction.

cs.CV↗

arXiv2026

General365: Benchmarking General Reasoning in Large Language Models Across Diverse and Challenging Tasks

Contemporary large language models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in specialized domains like mathematics and physics. However, their ability to generalize these reasoning skills to more general and broader contexts--often termed general reasoning--remains under-explored. Unlike domain-specific reasoning, general reasoning relies less on expert knowledge but still presents formidable reasoning challenges, such as complex constraints, nested logical branches, and semantic interference. To address this gap, we introduce General365, a benchmark specifically designed to assess general reasoning in LLMs. By restricting background knowledge to a K-12 level, General365 explicitly decouples reasoning from specialized expertise. The benchmark comprises 365 seed problems and 1,095 variant problems across eight categories, ensuring both high difficulty and diversity. Evaluations across 26 leading LLMs reveal that even the top-performing model achieves only 62.8% accuracy, in stark contrast to the near-perfect performances of LLMs in math and physics benchmarks. These results suggest that the reasoning abilities of current LLMs are heavily domain-dependent, leaving significant room for improvement in broader applications. We envision General365 as a catalyst for advancing LLM reasoning beyond domain-specific tasks toward robust, general-purpose real-world scenarios. Code, Dataset, and Leaderboard: https://general365.github.io

cs.CL↗

arXiv2026

Q-REAL: Towards Realism and Plausibility Evaluation for AI-Generated Content

Quality assessment of AI-generated content is crucial for evaluating model capability and guiding model optimization. However, most existing quality assessment datasets and models provide only a single quality score, which is too coarse to offer targeted guidance for improving generative models. In current applications of AI-generated images, realism and plausibility are two critical dimensions, and with the emergence of unified generation-understanding models, fine-grained evaluation along these dimensions becomes especially effective for improving generative performance. Therefore, we introduce Q-Real, a novel dataset for fine-grained evaluation of realism and plausibility in AI-generated images. Q-Real consists of 3,088 images generated by popular text-to-image models. For each image, we annotate the locations of major entities and provide a set of judgment questions and attribution descriptions for these along the dimensions of realism and plausibility. Considering that recent advances in multi-modal large language models (MLLMs) enable fine-grained evaluation of AI-generated images, we construct Q-Real Bench to evaluate them on two tasks: judgment and grounding with reasoning. Finally, to enhance MLLM capabilities, we design a fine-tuning framework and conduct experiments on multiple MLLMs using our dataset. Experimental results demonstrate the high quality and significance of our dataset and the comprehensiveness of the benchmark. Dataset and code will be released upon publication.

cs.CV↗

arXiv2026

TR-ICRL: Test-Time Rethinking for In-Context Reinforcement Learning

In-Context Reinforcement Learning (ICRL) enables Large Language Models (LLMs) to learn online from external rewards directly within the context window. However, a central challenge in ICRL is reward estimation, as models typically lack access to ground-truths during inference. To address this limitation, we propose Test-Time Rethinking for In-Context Reinforcement Learning (TR-ICRL), a novel ICRL framework designed for both reasoning and knowledge-intensive tasks. TR-ICRL operates by first retrieving the most relevant instances from an unlabeled evaluation set for a given query. During each ICRL iteration, LLM generates a set of candidate answers for every retrieved instance. Next, a pseudo-label is derived from this set through majority voting. This label then serves as a proxy to give reward messages and generate formative feedbacks, guiding LLM through iterative refinement. In the end, this synthesized contextual information is integrated with the original query to form a comprehensive prompt, with the answer determining through a final round of majority voting. TR-ICRL is evaluated on mainstream reasoning and knowledge-intensive tasks, where it demonstrates significant performance gains. Remarkably, TR-ICRL improves Qwen2.5-7B by 21.23% on average on MedQA and even 137.59% on AIME2024. Extensive ablation studies and analyses further validate the effectiveness and robustness of our approach. Our code is available at https://github.com/pangpang-xuan/TR_ICRL.

cs.CL↗