Elamur Research⌕
☰
home / papers / 2024 / Is LLM a Reliable Reviewer? A Comprehensive Evaluation of LLM on Automatic Paper Reviewing Tasks

Is LLM a Reliable Reviewer? A Comprehensive Evaluation of LLM on Automatic Paper Reviewing Tasks

Paper · 2024 · Ruiyang Zhou, Lu Chen and Kai Yu

Summary

The paper asks whether large language models can be trusted as automatic peer reviewers, a question earlier review datasets did not answer well because they did not verify that generated reviews were correct. The authors test GPT-3.5 (including the 16k-context variant) and GPT-4 on two existing task types. The first is aspect score prediction on PeerRead, where the model works either from a review or from parts of the paper. The second is review generation on ASAP, judged with automatic metrics, LLM-based scoring and manual checks. To measure correctness directly, they also build RR-MCQ, a set of multiple-choice questions taken from real ICLR-2023 review–rebuttal discussions. Each question is tagged by review aspect, paper part, required ability and need for outside information.

The models are useful but error-prone. They judge single answer options reasonably well, yet rarely answer a whole question correctly. They also struggle with long papers, zero-shot scoring and critical, technically detailed feedback. The study further shows that similarity metrics do not track true review quality, which is a caution for anyone building review-assistance tools.

Key points

  • RR-MCQ dataset: 196 questions derived from 55 reviews of 14 ICLR-2023 papers, released at huggingface.co/datasets/zhouruiyang/RR-MCQ. Labels cover:
    • review aspect (e.g., Soundness, Novelty, Reproducibility);
    • content aspect (e.g., Method, Dataset);
    • ability, ordered from Knowledge up to Correct;
    • whether extra information is needed.
  • RR-MCQ results: accuracy on individual options exceeds 60%, but fully correct answers occur only about 20% of the time.
  • Score prediction data: the ICLR-2017 subset of PeerRead, with about 1.3k manually annotated aspect scores (1–5, plus a "not discussed" label) over 427 reviews.
  • Score prediction results:
    • When given the review text, few-shot GPT-3.5 reaches an average Pearson correlation of about 0.65.
    • When given paper content instead, correlations drop to about 0.13–0.26, and the whole-paper GPT-3.5-16k setting performs worst.
  • Review generation: this task uses ASAP, and automatic similarity scores turn out to be poorly aligned with actual review quality.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksInformation Systems · 523 worksHistory and Philosophy of Science · 18 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 1,343 worksEvaluation & benchmarks · 685 works

Cite

Cite as Ruiyang Zhou, Lu Chen and Kai Yu. 2024. Is LLM a Reliable Reviewer? A Comprehensive Evaluation of LLM on Automatic Paper Reviewing Tasks. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{zhou2024is,
  title = {{Is LLM a Reliable Reviewer? A Comprehensive Evaluation of LLM on Automatic Paper Reviewing Tasks}},
  author = {Ruiyang Zhou, Lu Chen and Kai Yu},
  booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year = {2024},
  doi = {10.63317/48d359hjdvog},
  url = {https://doi.org/10.63317/48d359hjdvog}
}

References

27 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Liying Cheng, Lidong Bing, Qian Yu, Wei Lu, and Luo Si. 2020. Ape: Argument pair extraction from peer review and rebuttal via multi-task learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7000–7011. find
  2. Gautam Choudhary, Natwar Modani, and Nitish Maurya. 2021. React: A re view comment dataset for act ionability (and more). In Web Information Systems Engineering–WISE 2021: 22nd International Conference on Web Information Systems Engineering, WISE 2021, Melbourne, VIC, Australia, October 26–29, 2021,… find
  3. Mike D’Arcy, Alexis Ross, Erin Bransom, Bailey Kuehl, Jonathan Bragg, Tom Hope, and Doug Downey. 2023. Aries: A corpus of scientific paper edits made in response to peer reviews. arXiv preprint arXiv:2306.12587. find
  4. Nils Dycke, Ilia Kuznetsov, and Iryna Gurevych. 2022. Nlpeer: A unified resource for the computational study of peer review. arXiv preprint arXiv:2211.06651. find
  5. Michael Fromm, Evgeniy Faerman, Max Berrendorf, Siddharth Bhargava, Ruoxia Qi, Yao Zhang, Lukas Dennert, Sophia Selle, Yang Mao, and Thomas Seidl. 2021. Argument mining driven analysis of peer-reviews. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 4758–4766. find
  6. Tirthankar Ghosal, Sandeep Kumar, Prabhat Kumar Bharti, and Asif Ekbal. 2022. Peer review analyze: A novel benchmark resource for computational analysis of peer reviews. Plos one, 17(1):e0259238. find
  7. Xinyu Hua, Mitko Nikolov, Nikhil Badugu, and Lu Wang. 2019. Argument mining for understanding peer reviews. arXiv preprint arXiv:1903.10104. find
  8. Chao Jiang, Wei Xu, and Samuel Stevens. 2022. arxivedits: Understanding the human revision process in scientific writing. arXiv preprint arXiv:2210.15067. find
  9. Dongyeop Kang, Waleed Ammar, Bhavana Dalvi, Madeleine Van Zuylen, Sebastian Kohlmeier, Eduard Hovy, and Roy Schwartz. 2018. A dataset of peer reviews (peerread): Collection, insights and nlp applications. arXiv preprint arXiv:1804.09635.
  10. Neha Kennard, Tim O’Gorman, Rajarshi Das, Akshay Sharma, Chhandak Bagchi, Matthew Clinton, Pranay Kumar Yelugam, Hamed Zamani, and Andrew McCallum. 2021. Disapere: A dataset for discourse structure in peer review discussions. arXiv preprint arXiv:2110.08520. find
  11. Tom Kocmi and Christian Federmann. 2023. Large language models are state-of-the-art evaluators of translation quality. arXiv preprint arXiv:2302.14520. find
  12. Ilia Kuznetsov, Jan Buchmann, Max Eichler, and Iryna Gurevych. 2022. Revise and resubmit: An intertextual model of text-based collaboration in peer review. Computational Linguistics, 48(4):949–986. find
  13. Jiyi Li, Ayaka Sato, Kazuya Shimura, and Fumiyo Fukumoto. 2020. Multi-task peer-review score prediction. In Proceedings of the First Workshop on Scholarly Document Processing, pages 121– 126. find
  14. Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Yi Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Scott Smith, Yian Yin, et al. 2023. Can large language models provide useful feedback on research papers? a large-scale empirical analysis. arXiv preprint arXiv:2310.01783. find
  15. Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81.
  16. Jialiang Lin, Jiaxin Song, Zhangping Zhou, Yidong Chen, and Xiaodong Shi. 2023. Moprd: A multidisciplinary open peer review dataset. Neural Computing and Applications, pages 1–16. find
  17. Ryan Liu and Nihar B Shah. 2023. Reviewergpt? an exploratory study on using large language models for paper reviewing. arXiv preprint arXiv:2306.00622. find
  18. OpenAI. 2023. Gpt-4 technical report.
  19. Zachary Robertson. 2023. Gpt4 is slightly helpful for peer-review assistance: A pilot study. arXiv preprint arXiv:2307.05492. — [arXiv:2307.05492](https://arxiv.org/abs/2307.05492) find
  20. Chenhui Shen, Liying Cheng, Yang You, and Lidong Bing. 2023. Are large language models good evaluators for abstractive summarization? arXiv preprint arXiv:2305.13091. find
  21. Chenhui Shen, Liying Cheng, Ran Zhou, Lidong Bing, Yang You, and Luo Si. 2021. Mred: A meta-review dataset for structure-controllable text generation. arXiv preprint arXiv:2110.07474. find
  22. Shruti Singh, Mayank Singh, and Pawan Goyal. 2021. Compare: a taxonomy and dataset of comparison discussions in peer reviews. In 2021 ACM/IEEE Joint Conference on Digital Libraries (JCDL), pages 238–241. IEEE. find
  23. Oleg Vasilyev, Vedant Dharnidharka, and John Bohannon. 2020. Fill in the blanc: Human-free quality estimation of document summaries. arXiv preprint arXiv:2002.09836.
  24. Qingyun Wang, Qi Zeng, Lifu Huang, Kevin Knight, Heng Ji, and Nazneen Fatema Rajani. 2020. Reviewrobot: Explainable paper review generation based on knowledge synthesis. arXiv preprint arXiv:2010.06119. find
  25. Weizhe Yuan and Pengfei Liu. 2022. Kid-review: Knowledge-guided scientific review generation with oracle pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11639–11647. find
  26. Weizhe Yuan, Pengfei Liu, and Graham Neubig. 2022. Can we automate scientific reviewing? Journal of Artificial Intelligence Research, 75:171–212. find
  27. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2024-is-llm-reliable-reviewer-comprehensive/