PaperAudit-Bench: Benchmarking Error Detection in Research Papers for Critical Automated Peer Review

arXiv:2601.19916 · cs.CL · Submitted 2026-01-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "PaperAudit-Bench: Benchmarking Error Detection in Research Papers for Critical Automated Peer Review".

Jane: Large language models can generate fluent peer reviews, yet their assessments often lack sufficient critical rigor when substantive issues are subtle and distributed across a paper.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: The title of the paper is PaperAudit-Bench: Benchmarking Error Detection in Research Papers for Critical Automated Peer Review, and the authors are Songjun Tu, Yiwen Ma, Jiahao Lin, Qichao Zhang, Xiangyuan Lan, Junfeng Li, Nan Xu, and Linjing Li.

Jane: That sounds like a very specific system for testing the critical assessment abilities of AI review tools. It suggests they aren't just looking for surface-level mistakes in the writing.

Lu: The key idea is that existing models often generate fluent reviews but miss substantive issues that are hard to find when they are distributed across different parts of a paper, and this benchmark is designed to expose those weaknesses under long context settings.

Meng: So it's testing if an AI can spot something like a logical flaw in section one and then connect it to missing data in section three, which is what the authors call cross-section reasoning.

Lalam: The implication here is that we need benchmarks that specifically look for these kinds of distributed errors because they are where the real difficulty lies for current LLM review systems.

The paper's summary: Tom: The PaperAudit-Bench introduces two main things: first, the PaperAuditDataset, which is this error dataset covering both individual section errors and those needing cross-section reasoning under long context settings.

Jane: And the second part is the PaperAudit-Review framework, which integrates structured error detection with evidence-aware review generation to help support that critical assessment process.

Lu: They are using this dataset, which is built from papers accepted at top AI conferences like ICLR and ICML in two thousand twenty-five and they have about fifteen injected errors per paper on average.

Meng: The framework itself has three detection modes: Fast Mode for a quick scan, Standard Mode with explicit section awareness and a shared global memory, and Deep Mode which uses planning and multi-agent detection for specialized analysis.

Lalam: And the results show that using structured detection helps shift AI reviewers toward more critical evaluations grounded in identifiable technical issues, particularly with respect to technical quality and argumentative rigor.

The paper's improvements: Tom: They're not just presenting this idea; they are showing how they can train lightweight detectors using supervised fine-tuning and reinforcement learning on the PaperAuditDataset instead of needing massive large language model detectors.

Jane: Training these compact detectors involves SFT on structured data to map corrupted content to error descriptions, followed by RL using Group Relative Policy Optimization with a multi-dimensional reward function that captures accuracy, conciseness, and penalties for abnormal behavior.

Lu: This process yields compact detectors that have competitive detection performance while also improving output discipline compared to larger models.

Meng: From a deployment standpoint, having a lightweight detector is good because it means lower computational costs when you run the error detection part of the review pipeline.

Lalam: So the improvement lies in making these detectors efficient and disciplined, allowing us to integrate explicit error detection into the review workflow without requiring enormous model resources.

Conclusion: Tom: To wrap up, PaperAudit-Bench is presented as a unified benchmark for evaluating paper-level error detection and its interaction with automated peer review. The experiments show that structured detection changes review behavior in a more conservative and evidence-grounded manner.

Jane: So the main point is that this setup helps foster more critical and evidence-grounded automated peer review by using controlled benchmarks to see how these systems perform when they have to reason across the entire paper.

Lu: The work provides a reference point for future studies on designing error-aware automated reviewing systems, specifically focusing on how to expose distributed errors effectively.

Meng: I think the shift toward explicit error detection means that an AI reviewer is less likely to just be polite and more likely to actually challenge the technical claims based on detected flaws.

Lalam: I agree, and this whole PaperAudit-Bench setup gives researchers a concrete way to design these systems better moving forward.

Institute of Automation, Chinese Academy of Sciences. · University of Chinese Academy of Sciences. · Wenge Technology

cs.CL

Submitted: 2026-01-07

Updated: 2026-10-19

Journal ref: EMNLP 2026, Main

Code: https://github.com/TU2021/PaperAudit

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 71/100

The gist: Large language models can generate fluent peer reviews, yet their assessments often lack sufficient critical rigor when substantive issues are subtle and distributed across a paper.

Key concepts

PaperAuditDataset: This is a dataset of 220 research papers from top AI conferences (ICLR, ICML, NeurIPS 2025). Each paper is structured with section labels and contains about 15 injected errors. It's designed to be difficult by being long and having errors spread across different parts of the paper.
PaperAudit-Review Framework: This framework integrates error detection with review generation through three modes: Fast, Standard (using global memory for cross-section reasoning), and Deep (using multi-agent planning). The goal is to systematically shift reviewers toward more critical evaluations grounded in identifiable technical issues.
Lightweight Detectors Training: The system trains compact detectors using Supervised Fine-Tuning (SFT) on structured data, followed by Reinforcement Learning (RL) with Group Relative Policy Optimization (GRPO). This allows for training smaller models that achieve competitive detection performance while maintaining better efficiency and output discipline.
Evidence-Aware Review Generation: The review workflow uses the results from error detection to synthesize assessments. It moves beyond simple summarization by incorporating explicit error reports and novelty analysis, refining the baseline review under strict constraints to ensure critiques are strictly grounded in manuscript evidence.

Terminology

Benchmarking — 1× in this paper · explained in 4 episodes
Papers — 1× in this paper · explained in 3 episodes

Summary

Large language models can generate fluent peer reviews, yet their assessments often lack sufficient critical rigor when substantive issues are subtle and distributed across a paper.

PaperAudit-Bench

The gist is that PaperAudit-Bench introduces a comprehensive benchmark for auditing scientific papers and systematically evaluating the critical assessment behavior of automated peer review systems. This framework consists of two components: (1) PaperAuditDataset, an error dataset covering both errors identifiable within individual sections and those requiring cross-section reasoning, designed for controlled evaluation under long-context settings; and (2) PaperAudit-Review, an automated review framework that integrates structured error detection with evidence-aware review generation to support critical assessment.

PaperAuditDataset

The dataset is constructed from research papers accepted at the three top AI conferences: ICLR, ICML, and NeurIPS in 2025. The construction pipeline involves paper collection from OpenReview and filtering using a lightweight length-based criterion to exclude extremely long documents. Each paper is converted into a structured, section-aware JSON representation designed for multimodal language models, consisting of an ordered sequence of text and figure blocks, each annotated with a semantic section label. The dataset comprises 220 papers, with approximately 15 injected errors per paper on average. The dataset exhibits intrinsic properties such as covering a broad range of AI research subfields and being inherently difficult due to its long-context nature and sparse error distribution.

PaperAudit-Review Framework

The framework integrates structured error detection with evidence-aware review generation through two workflows. The detection workflow consists of three modes: (1) Fast Mode, which performs a single-pass scan over the entire paper; (2) Standard Mode, which introduces explicit section awareness and a shared global memory; and (3) Deep Mode, which incorporates explicit planning and specialized analysis via multi-agent detection. The review workflow builds upon the detection results to synthesize critical assessments.

Training Lightweight Detectors

The framework supports training lightweight detectors via SFT and RL to replace large-scale LLM detectors. The SFT stage trains the model on structured data derived from PaperAuditDataset, mapping corrupted content to structured error descriptions. The RL stage applies Group Relative Policy Optimization (GRPO) using a multi-dimensional reward function that captures detection accuracy, output conciseness, and abnormal behavior penalties. This process yields compact detectors with improved output discipline.

Application to Peer Review

Integrating explicit error detection into the review process systematically shifts AI reviewers toward more critical and discriminative evaluations grounded in identifiable technical and argumentative issues. The framework supports a detailed review workflow that begins with document understanding, followed by fine-grained error detection, novelty analysis via scholarly APIs, and finally, refinement of the baseline review using error and motivation reports. The system refines the baseline review under strict minimal-edit constraints to ensure that critiques are grounded exclusively in manuscript evidence.

Evaluation Results

Experiments reveal substantial variability in error detectability across models and detection depths, highlighting the difficulty of identifying such errors under long-context settings. Structured detection and document-level global memory are essential for exposing distributed errors and improving arXiv:2601.19916v1 [cs.CL] 7 Jan 2026 [3] The proposed dataset enables SFT and RL training of lightweight detectors, yielding compact models with competitive detection performance and improved output discipline.

Conclusion

The work presents PaperAudit-Bench as a unified benchmark for evaluating paper-level error detection and its interaction with automated peer review. Experiments demonstrate that structured detection changes review behavior in a more conservative and evidence-grounded manner. The goal is to foster more critical and evidence-grounded automated peer review through controlled benchmarks and structured error detection. The work provides a useful reference point for future studies on evaluating and designing error-aware automated reviewing systems.

Limitations

Key limitations include the use of synthetic errors, which cannot fully capture naturally occurring issues, and the potential for over-criticality when using deeper detection settings like Deep mode. Prompt and configuration sensitivity also affects performance trade-offs between coverage and strictness.

Ethical Considerations

This work uses LLMs solely for benchmark construction and evaluation; it does not encourage paper manipulation or advocate replacing human reviewers in real peer-review processes. LLM-based agents should be used at most for pre-submission self-auditing and review self-improvement, with all final judgments reserved for human experts.

References

Hui Chen, Miao Xiong, Yujie Lu, Wei Han, Ailin Deng, Yufei He, Jiaying Wu, Yibo Li, Yue Liu, and Bryan Hooi. 2025a. Mlr-bench: Evaluating ai agents on open-ended machine learning research. arXiv preprint arXiv:2505.19955 [1].

Qiguang Chen, Mingda Yang, Libo Qin, Jinhao Liu, Zheng Yan, Jiannan Guan, Dengyun Peng, Yiyan Ji, Hanjing Li, Mengkang Hu, and 1 others. 2025b. Ai4research: A survey of artificial intelligence for scientific research. arXiv preprint arXiv:2507.01903 [2].

Nils Dycke and Iryna Gurevych. 2025. Automatic reviewers fail to detect faulty reasoning in research papers: A new counterfactual evaluation framework. arXiv preprint arXiv:2508.21422 [3].

Yuqian Fu, Tinghong Chen, Jiajun Chai, Xihuai Wang, Songjun Tu, Guojun Yin, Wei Lin, Qichao Zhang, Yuanheng Zhu, and Dongbin Zhao. 2025. Srft: A single-stage method with supervised and reinforcement fine-tuning for reasoning. arXiv preprint arXiv:2506.19767 [4].

Aniketh Garikaparthi, Manasi Patwardhan, Lovekesh Vig, and Arman Cohan. 2025. Iris: Interactive research ideation system for accelerating scientific discovery. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 592–603 [10].

Junyi Hou, Andre Lin Huikai, Nuo Chen, Yiwei Gong, and Bingsheng He. 2025. Paperdebugger: A plugin-based multi-agent system for in-editor academic writing, review, and editing. arXiv preprint arXiv:2512.02589 [11].

Xiang Hu, Hongyu Fu, Jinge Wang, Yifeng Wang, Zhikun Li, Renjun Xu, Yu Lu, Yaochu Jin, Lili Pan, and Zhenzhong Lan. 2025. Nova: An iterative planning framework for enhancing scientific innovation with large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 21330–21359 [14].

Yiqiao Jin, Qinlin Zhao, Yiyang Wang, Hao Chen, Kaijie Zhu, Yijia Xiao, and Jindong Wang. 2024. Agentreview: Exploring peer review dynamics with llm agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, page 1208–1226 [15].

Patrick Tser Jern Kon, Jiachen Liu, Xinyi Zhu, Qiuyi Ding, Jingjia Peng, Jiarong Xing, Yibo Huang, Yiming Qiu, Jayanth Srinivasa, Myungjin Lee, and 1 others. 2025. Exp-bench: Can ai conduct ai research experiments? arXiv preprint arXiv:2505.24785 [16].

Jiatao Li, Yanheng Li, Xinyu Hu, Mingqi Gao, and Xiaojun Wan. 2025a. Where do llms go wrong? diagnosing automated peer review via aspect-guided multi-level perturbation. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 1572–1581 [17].

Long Li, Weiwen Xu, Jiayan Guo, Ruochen Zhao, Xingxuan Li, Yuqian Yuan, Boqiang Zhang, Yuming Jiang, Yifei Xin, Ronghao Dang, and 1 others. 2025b. Chain of ideas: Revolutionizing research via novel idea development with llm agents. Findings of the Association for Computational Linguistics: EMNLP 2025, pages 8971–9004 [18].

Ruochi Li, Haoxuan Zhang, Edward Gehringer, Ting Xiao, Junhua Ding, and Haihua Chen. 2025c. Unveiling the merits and defects of llms in automatic review generation for scientific papers.

Improvements for AI systems

System-driven error detection will move beyond surface-level checks to systematically probe for distributed errors by utilizing three distinct modes: Fast, Standard with document-level global memory, and Deep with plan-driven multi-agent detection.
The improved system can perform cross-section reasoning by leveraging the Standard Mode, which uses a shared global memory to identify inconsistencies that are distributed across distant parts of the paper.
A lightweight detector can be trained via SFT and RL on the PaperAuditDataset, enabling effective error detection at reduced computational cost, as demonstrated by models achieving performance comparable to larger general-purpose models while maintaining a better EC–FP trade-off.
The automated review framework will produce systematically stricter and more discriminative evaluations by integrating explicit error detection into the workflow, shifting AI reviewers toward more critical and discriminative evaluations grounded in identifiable technical and argumentative issues.

Abstract

Large language models (LLMs) can generate fluent peer reviews but often miss sparse, subtle, and document-level scientific errors. We introduce PaperAudit-Bench, a benchmark for evaluating paper-level error detection and error-aware automated review. It comprises (1) PaperAudit-Dataset, 220 high-quality AI/ML papers containing more than 2,200 controlled synthetic errors across eight categories, and (2) PaperAudit-Review, an evidence-aware framework that links structured error detection to review generation. Experiments across models, error categories, paper sections, and analysis depths show that sparse long-context auditing remains challenging for current LLMs. Conditioning review generation on explicit, manuscript-grounded error evidence produces more conservative and discriminative assessments that align more closely with human judgments than representative automated-review baselines. The dataset also supports post-training compact detectors through SFT and RL; Qwen3-14B achieves 55.8% F1 at Detection@1 under a fixed held-out protocol. A PubPeer transfer study and deployment in ScienceOne, a scientific foundation-model platform led by the Chinese Academy of Sciences, further illustrate practical applicability. Code and Dataset Page: https://github.com/TU2021/PaperAudit ScienceOne Manuscript Review: https://www.scienceone.cn/lit/#/assistant