Searcharxiv⌕ Search

arXiv subjects

Cunxiang Wang

Publications and source records attributed to Cunxiang Wang.

At least 19 recordsLinked to original sources
arXiv2026

MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding

Recently, the rapid development of large language models (LLMs) has reshaped software engineering by enabling autonomous code agents that plan, execute, and utilize external tools iteratively to tackle complex tasks. Beyond achieving functional correctness, these agents must faithfully follow process instructions and constraints throughout the development lifecycle. However, existing benchmarks typically focus on final functional correctness or confine instruction-following evaluation to single-turn, general chat or simple code generation scenarios, leaving instruction-following in multi-turn agentic coding underexplored. To bridge this gap, we propose MTAC-IFBench, a comprehensive benchmark for this critical capability. It features multi-turn progressive software development instructions with diverse constraints spanning 6 primary and 18 secondary categories. With an average of 7.04 turns and 91.33 constraints per instance, it poses a rigorous challenge to current LLMs. To make the evaluation reliable, we construct a checklist for each constraint and functional requirement, and integrate verification scripts and judge agents to verify each checklist item. MTAC-IFBench identifies significant deficiencies in existing code agents in multi-turn instruction-following, with their performance degrading rapidly as the interaction session grows longer.

cs.CL↗
arXiv2026

From Base Rollouts to RL Reasoning: A Budgeted Search Perspective

Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and search remains unclear. Does RL create reasoning the base model lacks, or shift the rollout distribution toward trajectories it can already reach but rarely samples? We study this behaviorally with a Unified Decoding Framework (UDF), which expresses token-level sampling, beam-like search, tree search, and sequence-level resampling as executable policies over a shared budgeted operating space, scored post hoc with pass@$k$, self-consistency, best-of-$N$, and first-finish success. Using paired Base/RL checkpoints from SimpleRL-Zoo, we ask whether an RL default-policy curve can be approximated by a structured path of Base operating points. On Math500, AIME, GPQA, and IFEval, the pass@$k$ recovery path follows a Budgeted Operating-Point Transition Rule (BOPTR), $N_{\mathrm{Base}} \approx αN_{\mathrm{RL}}^β$, with benchmark-conditioned exponents. On Qwen2.5-7B, BOPTR gives the lowest transfer error among the non-oracle rules we test, 3.41 pp (95% CI [2.32, 5.53]); a three-seed replication gives 3.07 $\pm$ 0.39 pp. The rule extends to ten models across four families (3.28 to 4.87 pp on checkpoints added after fitting), to four benchmarks it was never fitted on (5.03 pp vs. 4.44 pp in fit), and holds without an RL checkpoint for the target model (4.19 pp) or without RL supervision of any kind (5.08 pp). These results support a qualified internalized-search reading: under the recipe we test, much of the measured RL gain corresponds to a change in sampling efficiency toward operating points the base model can already reach under search. We treat the scaling patterns as descriptive of this recipe and cohort, report where they break down, and use UDF and BOPTR as behavioral diagnostics rather than evidence of parameter-level equivalence.

cs.CL↗
arXiv2026

HoWToBench: Holistic Evaluation for LLM's Capability in Human-level Writing using Tree of Writing

Evaluating the writing capabilities of large language models (LLMs) remains a significant challenge due to the multidimensional nature of writing skills and the limitations of existing metrics. LLM's performance in thousand-words level and open-ended writing is inadequately assessed by traditional reference-based metrics or modern LLM-as-a-judge methods. We propose Tree-of-Writing (ToW), to resolve the implicit inconsistency often found when LLM-as-a-judge aggregates all sub-features in text evaluation. ToW incorporates a tree-structured workflow by explicitly modeling the aggregation weights of sub-features. We also present HowToBench, a large-scale Chinese writing benchmark encompassing 12 genres and 1302 instructions across three task categories: contextual completion, outline-guided writing, and open-ended generation. ToW successfully mitigates the biases, achieving a 0.93 Pearson correlation with human judgments. Furthermore, we detect that both overlap-based text generation metrics and popular LLM-as-a-judge practices are vulnerable to textual disturbances, while ToW is robust to them. We also uncover a negative correlation between input length and content-related scores in the Guide task, showcasing that it cannot be simply improved by input-side information piling.

cs.CL↗
arXiv2026

IF-CRITIC: Towards a Fine-Grained LLM Critic for Instruction-Following Evaluation

Instruction-following is a fundamental ability of Large Language Models (LLMs), requiring their generated outputs to follow multiple constraints imposed in input instructions. Numerous studies have attempted to enhance this ability through preference optimization or reinforcement learning based on reward signals from LLM-as-a-Judge. However, existing evaluation models for instruction-following still possess many deficiencies, such as substantial costs and unreliable assessments. To this end, we propose IF-CRITIC, an LLM critic for fine-grained, efficient, and reliable instruction-following evaluation. We first develop a checklist generator to decompose instructions and generate constraint checklists. With the assistance of the checklists, we collect high-quality critique training data through a multi-stage critique filtering mechanism and employ a constraint-level preference optimization method to train IF-CRITIC. Extensive experiments show that the evaluation performance of IF-CRITIC can beat strong LLM-as-a-Judge baselines, including o4-mini and Gemini-3-Pro. With the reward signals provided by IF-CRITIC, LLMs can achieve substantial performance gains in instruction-following optimization under lower computational overhead compared to strong LLM critic baselines. Our code and model are available at https://github.com/thu-coai/IF-CRITIC.

cs.CL↗
arXiv2026

Beyond Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation

Large Language Models (LLMs) have significantly advanced Machine Translation (MT), applying them to linguistically complex domains-such as Social Network Services, literature etc. In these scenarios, translations often require handling non-literal expressions, leading to the inaccuracy of MT metrics. To systematically investigate the reliability of MT metrics, we first curate a meta-evaluation dataset focused on non-literal translations, namely MENT. MENT encompasses four non-literal translation domains and features source sentences paired with translations from diverse MT systems, with 7,530 human-annotated scores on translation quality. Experimental results reveal the inaccuracies of traditional MT metrics and the limitations of LLM-as-a-Judge, particularly the knowledge cutoff and score inconsistency problem. To mitigate these limitations, we propose RATE, a novel agentic translation evaluation framework, centered by a reflective Core Agent that dynamically invokes specialized sub-agents. Experimental results indicate the efficacy of RATE, achieving an improvement of at least 3.2 points in combined system- and segment-level correlation with human judgments compared with current methods. Further experiments demonstrate the robustness of RATE to general-domain MT evaluation. Code and dataset are available at: https://github.com/BITHLP/RATE.

cs.CL↗
arXiv2026

IF-RewardBench: Benchmarking Judge Models for Instruction-Following Evaluation

Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.

cs.CL↗
arXiv2026

IndustryCode: A Benchmark for Industry Code Generation

Code generation and comprehension by Large Language Models (LLMs) have emerged as core drivers of industrial intelligence and decision optimization, finding widespread application in fields such as finance, automation, and aerospace. Although recent advancements have demonstrated the remarkable potential of LLMs in general code generation, existing benchmarks are mainly confined to single domains and languages. Consequently, they fail to effectively evaluate the generalization capabilities required for real-world industrial applications or to reflect the coding proficiency demanded by complex industrial scenarios. To bridge this gap, we introduce IndustryCode, the first comprehensive benchmark designed to span multiple industrial domains and programming languages. IndustryCode comprises 579 sub-problems derived from 125 primary industrial challenges, accompanied by rigorous problem descriptions and test cases. It covers a wide range of fields, including finance, automation, aerospace, and remote sensing-and incorporates diverse programming languages such as MATLAB, Python, C++, and Stata. In our evaluation, the top-performing model, Claude 4.5 Opus, achieved an overall accuracy of 68.1% on sub-problems and 42.5% main problems. The benchmark dataset and automated evaluation code will be made publicly available upon acceptance.

cs.SE↗
arXiv2026

MVSS: A Unified Framework for Multi-View Structured Survey Generation

Scientific surveys require not only summarizing large bodies of literature, but also organizing them into clear and coherent conceptual structures. However, existing automatic survey generation methods typically focus on linear text generation and struggle to explicitly model hierarchical relations among research topics and structured methodological comparisons, resulting in substantial gaps in structural organization and evidence presentation compared to expert-written surveys. To address this limitation, we propose MVSS, a multi-view structured survey generation framework that jointly generates and aligns citation-grounded hierarchical trees, structured comparison tables, and survey text. MVSS follows a structure-first paradigm: it first constructs a tree that captures the conceptual organization of a research domain, then generates comparison tables constrained by the tree structure, and finally uses both the tree and tables as joint structural constraints to guide outline construction and survey text generation. This design enables complementary and aligned multi-view representations across structure, comparison, and narrative. In addition, we introduce a dedicated evaluation framework that systematically assesses generated surveys from multiple dimensions, including structural quality, comparative completeness, and citation fidelity. Through large-scale experiments on 76 computer science topics, we demonstrate that MVSS significantly outperforms existing methods in survey organization and evidence grounding, and achieves performance comparable to expert-written surveys across multiple evaluation metrics.

cs.CL↗
arXiv2026

TraceSIR: A Multi-Agent Framework for Structured Analysis and Reporting of Agentic Execution Traces

Agentic systems augment large language models with external tools and iterative decision making, enabling complex tasks such as deep research, function calling, and coding. However, their long and intricate execution traces make failure diagnosis and root cause analysis extremely challenging. Manual inspection does not scale, while directly applying LLMs to raw traces is hindered by input length limits and unreliable reasoning. Focusing solely on final task outcomes further discards critical behavioral information required for accurate issue localization. To address these issues, we propose TraceSIR, a multi-agent framework for structured analysis and reporting of agentic execution traces. TraceSIR coordinates three specialized agents: (1) StructureAgent, which introduces a novel abstraction format, TraceFormat, to compress execution traces while preserving essential behavioral information; (2) InsightAgent, which performs fine-grained diagnosis including issue localization, root cause analysis, and optimization suggestions; (3) ReportAgent, which aggregates insights across task instances and generates comprehensive analysis reports. To evaluate TraceSIR, we construct TraceBench, covering three real-world agentic scenarios, and introduce ReportEval, an evaluation protocol for assessing the quality and usability of analysis reports aligned with industry needs. Experiments show that TraceSIR consistently produces coherent, informative, and actionable reports, significantly outperforming existing approaches across all evaluation dimensions. Our project and video are publicly available at https://github.com/SHU-XUN/TraceSIR.

cs.AI↗
arXiv2026

RAVEL: Reasoning Agents for Validating and Evaluating LLM Text Synthesis

Large Language Models have evolved from single-round generators into long-horizon agents, capable of complex text synthesis scenarios. However, current evaluation frameworks lack the ability to assess the actual synthesis operations, such as outlining, drafting, and editing. Consequently, they fail to evaluate the actual and detailed capabilities of LLMs. To bridge this gap, we introduce RAVEL, an agentic framework that enables the LLM testers to autonomously plan and execute typical synthesis operations, including outlining, drafting, reviewing, and refining. Complementing this framework, we present C3EBench, a comprehensive benchmark comprising 1,258 samples derived from professional human writings. We utilize a "reverse-engineering" pipeline to isolate specific capabilities across four tasks: Cloze, Edit, Expand, and End-to-End. Through our analysis of 14 LLMs, we uncover that most LLMs struggle with tasks that demand contextual understanding under limited or under-specified instructions. By augmenting RAVEL with SOTA LLMs as operators, we find that such agentic text synthesis is dominated by the LLM's reasoning capability rather than raw generative capacity. Furthermore, we find that a strong reasoner can guide a weaker generator to yield higher-quality results, whereas the inverse does not hold. Our code and data are available at this link: https://github.com/ZhuoerFeng/RAVEL-Reasoning-Agents-Text-Eval.

cs.CL↗
arXiv2026

RLAR: An Agentic Reward System for Multi-task Reinforcement Learning on Large Language Models

Large language model alignment via reinforcement learning depends critically on reward function quality. However, static, domain-specific reward models are often costly to train and exhibit poor generalization in out-of-distribution scenarios encountered during RL iterations. We present RLAR (Reinforcement Learning from Agent Rewards), an agent-driven framework that dynamically assigns tailored reward functions to individual queries. Specifically, RLAR transforms reward acquisition into a dynamic tool synthesis and invocation task. It leverages LLM agents to autonomously retrieve optimal reward models from the Internet and synthesize programmatic verifiers through code generation. This allows the reward system to self-evolve with the shifting data distributions during training. Experimental results demonstrate that RLAR yields consistent performance gains ranging from 10 to 60 across mathematics, coding, translation, and dialogue tasks. On RewardBench-V2, RLAR significantly outperforms static baselines and approaches the performance upper bound, demonstrating superior generalization through dynamic reward orchestration. The data and code are available on this link: https://github.com/ZhuoerFeng/RLAR.

cs.CL↗
arXiv2026

GLM-5: from Vibe Coding to Agentic Engineering

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

cs.LG↗
arXiv2026

DVD: A Robust Method for Detecting Variant Contamination in Large Language Model Evaluation

Evaluating large language models (LLMs) is increasingly confounded by \emph{variant contamination}: the training corpus contains semantically equivalent yet lexically or syntactically altered versions of test items. Unlike verbatim leakage, these paraphrased or structurally transformed variants evade existing detectors based on sampling consistency or perplexity, thereby inflating benchmark scores via memorization rather than genuine reasoning. We formalize this problem and introduce \textbf{DVD} (\textbf{D}etection via \textbf{V}ariance of generation \textbf{D}istribution), a single-sample detector that models the local output distribution induced by temperature sampling. Our key insight is that contaminated items trigger alternation between a \emph{memory-adherence} state and a \emph{perturbation-drift} state, yielding abnormally high variance in the synthetic difficulty of low-probability tokens; uncontaminated items remain in drift with comparatively smooth variance. We construct the first benchmark for variant contamination across two domains Omni-MATH and SuperGPQA by generating and filtering semantically equivalent variants, and simulate contamination via fine-tuning models of different scales and architectures (Qwen2.5 and Llama3.1). Across datasets and models, \textbf{DVD} consistently outperforms perplexity-based, Min-$k$\%++, edit-distance (CDD), and embedding-similarity baselines, while exhibiting strong robustness to hyperparameters. Our results establish variance of the generation distribution as a principled and practical fingerprint for detecting variant contamination in LLM evaluation.

cs.AI↗
arXiv2025

Deep Research: A Systematic Survey

Large language models (LLMs) have rapidly evolved from text generators into powerful problem solvers. Yet, many open tasks demand critical thinking, multi-source, and verifiable outputs, which are beyond single-shot prompting or standard retrieval-augmented generation. Recently, numerous studies have explored Deep Research (DR), which aims to combine the reasoning capabilities of LLMs with external tools, such as search engines, thereby empowering LLMs to act as research agents capable of completing complex, open-ended tasks. This survey presents a comprehensive and systematic overview of deep research systems, including a clear roadmap, foundational components, practical implementation techniques, important challenges, and future directions. Specifically, our main contributions are as follows: (i) we formalize a three-stage roadmap and distinguish deep research from related paradigms; (ii) we introduce four key components: query planning, information acquisition, memory management, and answer generation, each paired with fine-grained sub-taxonomies; (iii) we summarize optimization techniques, including prompting, supervised fine-tuning, and agentic reinforcement learning; and (iv) we consolidate evaluation criteria and open challenges, aiming to guide and facilitate future development. As the field of deep research continues to evolve rapidly, we are committed to continuously updating this survey to reflect the latest progress in this area.

cs.CL↗
arXiv2025

UDA: Unsupervised Debiasing Alignment for Pair-wise LLM-as-a-Judge

Pairwise evaluation of Large Language Models (LLMs) is a common paradigm, but it is prone to preference bias, where judges systematically favor certain outputs, such as their own. This bias leads to inconsistent and skewed rankings across different judges. To address this, we first empirically demonstrate significant and heterogeneous biases in cross-model evaluations. We then propose UDA (Unsupervised Debiasing Alignment), a framework that reduces inter-judge disagreement by dynamically adjusting the Elo rating system. For each pairwise comparison, a compact neural network learns to adaptively set the K-factor and refine win probabilities. Crucially, UDA operates in a fully unsupervised manner, guided solely by the objective of minimizing the dispersion among the Elo trajectories of all judges. This forces an alignment towards a collective consensus, which serves as an unsupervised proxy for a more stable and reproducible evaluation. In addition, we provide theoretical motivation demonstrating how alignment towards a consensus can reduce aggregate system bias. Experiments show that UDA significantly reduces the inter-judge rating standard deviation by up to 63.4% and improves the average correlation with human judgments by 24.7%. Notably, UDA elevates the performance of poorly performing judges to achieve parity with high-quality ones, fostering a more robust and reliable evaluation ecosystem. Code and data are available at https://anonymous.4open.science/r/62AB93CD-23B4.

cs.AI↗
arXiv2025

Deep Literature Survey Automation with an Iterative Workflow

Automatic literature survey generation has attracted increasing attention, yet most existing systems follow a one-shot paradigm, where a large set of papers is retrieved at once and a static outline is generated before drafting. This design often leads to noisy retrieval, fragmented structures, and context overload, ultimately limiting survey quality. Inspired by the iterative reading process of human researchers, we propose \ours, a framework based on recurrent outline generation, in which a planning agent incrementally retrieves, reads, and updates the outline to ensure both exploration and coherence. To provide faithful paper-level grounding, we design paper cards that distill each paper into its contributions, methods, and findings, and introduce a review-and-refine loop with visualization enhancement to improve textual flow and integrate multimodal elements such as figures and tables. Experiments on both established and emerging topics show that \ours\ substantially outperforms state-of-the-art baselines in content coverage, structural coherence, and citation quality, while producing more accessible and better-organized surveys. To provide a more reliable assessment of such improvements, we further introduce Survey-Arena, a pairwise benchmark that complements absolute scoring and more clearly positions machine-generated surveys relative to human-written ones. The code is available at https://github.com/HancCui/IterSurvey\_Autosurveyv2.

cs.CL↗
arXiv2025

Rethinking LLM Evaluation: Can We Evaluate LLMs with 200x Less Data?

As the demand for comprehensive evaluations of diverse model capabilities steadily increases, benchmark suites have correspondingly grown significantly in scale. Despite notable advances in redundancy reduction and subset-level performance prediction, a systematic framework that effectively integrates these methods to ensure both prediction accuracy and ranking consistency is still largely elusive. In this paper, we first perform a sample-level analysis of benchmark redundancy and identify several highly similar samples that can be eliminated. Besides, we frame benchmark compression as an optimization problem with the aim of score reconstruction. Building on these, we then propose EssenceBench, a coarse-to-fine framework utilizing an iterative Genetic Algorithm (GA), which takes the advantages of fitness-based subset search and attribution-based sample search. Compared to previous methods, our approach yields superior compression results with lower reconstruction error and markedly higher efficiency. In particular, on the HellaSwag benchmark (10K samples), our method preserves the ranking of all models shifting within 5% using 25x fewer samples, and achieves 95% ranking preservation shifting within 5% using only 200x fewer samples.

cs.CL↗
arXiv2025

AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs

Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.

cs.SD↗