Elamur Research⌕
☰
home / papers / 2023 / Can large language models provide useful feedback on research papers? A large-scale empirical analysis

Can large language models provide useful feedback on research papers? A large-scale empirical analysis

Paper · 2023 · Weixin Liang, Yuhui Zhang, Hancheng Cao et al.

Summary

Rising submission volumes make it harder to get timely, high-quality critique of manuscripts. ICLR submissions, for example, grew from 960 in 2018 to 4,966 in 2023. Junior and under-resourced researchers are hit hardest. The authors test whether GPT-4 can supply useful reviewer-style comments. Their pipeline parses a paper's full PDF and packs the title, abstract, captions and body text into a structured prompt. In a single pass, GPT-4 then comments on significance and novelty, reasons to accept, reasons to reject, and possible improvements.

Quality was assessed in two ways. Retrospectively, comments from GPT-4 and from real reviewers (15 Nature family journals and ICLR) were automatically extracted and semantically matched. The share of points GPT-4 had in common with a reviewer was about the same as the share two human reviewers had in common. Prospectively, researchers rated GPT-4 feedback on their own papers, and most found it useful. However, the model favours generic requests and seldom critiques method design in depth. LLM feedback therefore looks like a complement to expert review, most valuable for early drafts.

Key points

  • Data: 3,096 accepted Nature family papers (8,745 reviewer comments) and 1,709 ICLR 2022–2023 papers, both accepted and rejected (6,505 comments). Only post-2022 papers were used, to avoid overlap with GPT-4's training data (cutoff Sep 2021).
  • Matching pipeline: Human checks gave F1 scores of 96.8% for comment extraction and 82.4% for matching.
  • Overlap with individual reviewers: GPT-4 vs. one reviewer was 30.85% on Nature data (human–human: 28.58%) and 39.23% on ICLR (human–human: 35.25%). For rejected ICLR papers it rose to 43.80%. On Nature data, 57.55% of GPT-4's points were raised by at least one reviewer, and a paper-shuffling control showed the feedback is not merely generic.
  • User study: 308 researchers from 110 US institutions in AI and computational biology took part. 57.4% rated the feedback helpful or very helpful, and 82.4% preferred it to the input of at least some human reviewers.
  • Code: github.com/Weixin-Liang/LLM-scientific-feedback

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

 Institutions

🎓Stanford UniversityStanford · United States506 works · 330,052 citations🎓Cornell UniversityIthaca · United States134 works · 65,159 citations🎓Northwestern UniversityEvanston · United States46 works · 2,850 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksDecision Sciences · 249 worksStatistics, Probability and Uncertainty · 65 worksHistory and Philosophy of Science · 62 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Smith, Yian Yin, Daniel McFarland and James Zou. 2023. Can large language models provide useful feedback on research papers? A large-scale empirical analysis. arXiv:2310.01783.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2310.01783 · arXiv 2310.01783

@misc{liang2023can,
  title = {{Can large language models provide useful feedback on research papers? A large-scale empirical analysis}},
  author = {Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Smith, Yian Yin, Daniel McFarland and James Zou},
  year = {2023},
  doi = {10.48550/arxiv.2310.01783},
  eprint = {2310.01783},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2310.01783}
}

References

61 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shannon, C. E. A mathematical theory of communication. The Bell system technical journal 27, 379–423 (1948). find
  2. Tribus, M. & McIrvine, E. C. Energy and information. Sci. Am. 225, 179–190 (1971). find
  3. Kuhn, T. S. The structure of scientifi revolutions. The Un Chic. Press. 2, 90 (1962). find
  4. Horbach, S. P. & Halffman, W. The changing forms and expectations of peer review. Res. integrity peer review 3, 1–15 (2018). find
  5. Price, D. J. D. S. Little science, big science (Columbia University Press, 1963). find
  6. Jones, B. F. The burden of knowledge and the “death of the renaissance man”: Is innovation getting harder? The Rev. Econ. Stud. 76, 283–317 (2009). find
  7. Aczel, B., Szaszi, B. & Holcombe, A. O. A billion-dollar donation: estimating the cost of researchers’ time spent on peer review. Res. Integr. Peer Rev. 6, 1–8 (2021). find
  8. Alberts, B., Hanson, B. & Kelner, K. L. Reviewing peer review (2008). find
  9. Björk, B.-C. & Solomon, D. The publishing delay in scholarly peer-reviewed journals. J. informetrics 7, 914–923 (2013). find
  10. Lee, C. J., Sugimoto, C. R., Zhang, G. & Cronin, B. Bias in peer review. J. Am. Soc. for information Sci. Technol. 64, 2–17 (2013). find
  11. Kovanis, M., Porcher, R., Ravaud, P. & Trinquart, L. The global burden of journal peer review in the biomedical literature: Strong imbalance in the collective enterprise. PloS one 11, e0166387 (2016). find
  12. Shah, N. B. Challenges, experiments, and computational solutions in peer review. Commun. ACM 65, 76–87 (2022). find
  13. Bourdieu, P. Cultural reproduction and social reproduction. In Knowledge, education, and cultural change, 71–112 (Routledge, 2018). find
  14. Merton, R. K. The matthew effect in science: The reward and communication systems of science are considered. Science 159, 56–63 (1968). find
  15. Chu, J. S. & Evans, J. A. Slowed canonical progress in large fields of science. Proc. Natl. Acad. Sci. 118, e2021636118 (2021). find
  16. Bloom, N., Jones, C. I., Van Reenen, J. & Webb, M. Are ideas getting harder to find? Am. Econ. Rev. 110, 1104–1144 (2020). find
  17. Brown, T. et al. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. 33, 1877–1901 (2020).
  18. Ouyang, L. et al. Training language models to follow instructions with human feedback. Adv. Neural Inf. Process. Syst. 35, 27730–27744 (2022).
  19. OpenAI. GPT-4 Technical Report. ArXiv abs/2303.08774 (2023).
  20. Ayers, J. W. et al. Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA internal medicine (2023). find
  21. Lee, M. et al. Evaluating human-language model interaction. arXiv preprint arXiv:2212.09746 (2022). find
  22. Kung, T. H. et al. Performance of chatgpt on usmle: Potential for ai-assisted medical education using large language models. PLoS digital health 2, e0000198 (2023).
  23. Terwiesch, C. Would chat gpt3 get a wharton mba? a prediction based on its performance in the operations management course. Mack Inst. for Innov. Manag. at Whart. Sch. Univ. Pennsylvania (2023). find
  24. Schulz, R. et al. Is the future of peer review automated? BMC Res. Notes 15, 1–5 (2022). find
  25. Liu, R. & Shah, N. B. Reviewergpt? an exploratory study on using large language models for paper reviewing. arXiv preprint arXiv:2306.00622 (2023).
  26. Robertson, Z. Gpt4 is slightly helpful for peer-review assistance: A pilot study. arXiv preprint arXiv:2307.05492 (2023). — [arXiv:2307.05492](https://arxiv.org/abs/2307.05492) find
  27. Nature. How to Write a Report. https://www.nature.com/nature/for-referees/how-to-write-a-report. Accessed: 21 September 2023. ↗
  28. Nature Communications. Writing Your Report. writing-your-report. Accessed: 21 September 2023. find
  29. Rogers, A. & Augenstein, I. How to review for ACL Rolling Review. https://aclrollingreview.org/reviewertutorial (2021). ↗
  30. Association for Computational Linguistics. ACL’23 Peer Review Policies. https://2023.aclweb.org/blog/review-acl23/ (2023). ↗
  31. Association for Computational Linguistics. ACL-IJCNLP 2021 Instructions for Reviewers. https://2021.aclweb.org/blog/instructions-for-reviewers/ (2021). ↗
  32. International Conference on Machine Learning. ICML 2023 Reviewer Tutorial. https://icml.cc/Conferences/2023/ReviewerTutorial (2023). ↗
  33. Nicholas, K. A. & Gordon, W. S. A quick guide to writing a solid peer review. Eos, Transactions Am. Geophys. Union 92, 233–234 (2011). find
  34. Luhn, H. P. The automatic creation of literature abstracts. IBM J. research development 2, 159–165 (1958). find
  35. Edmundson, H. P. New methods in automatic extracting. J. ACM (JACM) 16, 264–285 (1969). find
  36. Mihalcea, R. & Tarau, P. Textrank: Bringing order into text. In Proceedings of the 2004 conference on empirical methods in natural language processing, 404–411 (2004). find
  37. Erkan, G. & Radev, D. R. Lexrank: Graph-based lexical centrality as salience in text summarization. J. artificial intelligence research 22, 457–479 (2004). find
  38. Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer, T. K. & Harshman, R. Indexing by latent semantic analysis. J. Am. society for information science 41, 391–407 (1990). https://www.nature.com/ncomms/for-reviewers/ ↗
  39. Socher, R., Huang, E., Pennin, J., Manning, C. D. & Ng, A. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. Adv. neural information processing systems 24 (2011). find
  40. Bowman, S. R., Angeli, G., Potts, C. & Manning, C. D. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326 (2015).
  41. Birhane, A. et al. The values encoded in machine learning research. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’22, 173–184, DOI: 10.1145/3531146.3533083 (Association for Computing Machinery, New York, NY, USA, 2022). — [doi:10.1145/3531146.3533083](https://doi.org/10.1145/3531146.3533083) find
  42. Smith, J. J., Amershi, S., Barocas, S., Wallach, H. & Wortman Vaughan, J. Real ml: Recognizing, exploring, and articulating limitations of machine learning research. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 587–597 (2022). find
  43. Koch, B., Denton, E., Hanna, A. & Foster, J. G. Reduced, reused and recycled: The life of a dataset in machine learning research. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021). find
  44. Scheuerman, M. K., Hanna, A. & Denton, E. Do datasets have politics? disciplinary values in computer vision dataset development. Proc. ACM on Human-Computer Interact. 5, 1–37 (2021). find
  45. Meyer, B. D., Mok, W. K. & Sullivan, J. X. Household surveys in crisis. J. Econ. Perspectives 29, 199–226 (2015). find
  46. Ross, M. B. et al. Women are credited less in science than men. Nature 608, 135–145 (2022). find
  47. Noy, S. & Zhang, W. Experimental evidence on the productivity effects of generative artificial intelligence. Available at SSRN 4375283 (2023). find
  48. Peng, S., Kalliamvakou, E., Cihon, P. & Demirer, M. The impact of ai on developer productivity: Evidence from github copilot. arXiv preprint arXiv:2302.06590 (2023). — [arXiv:2302.06590](https://arxiv.org/abs/2302.06590) find
  49. Epstein, Z. et al. Art and the science of generative ai. Science 380, 1110–1111 (2023). find
  50. Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023). find
  51. Nature will publish peer review reports as a trial. Nature 578, 8, DOI: 10.1038/d41586-020-00309-9 (2020). — [doi:10.1038/d41586-020-00309-9](https://doi.org/10.1038/d41586-020-00309-9) find
  52. Nature is trialling transparent peer review - the early results are encouraging. Nature 603, 8, DOI: 10.1038/d41586-022-00493-w (2022). — [doi:10.1038/d41586-022-00493-w](https://doi.org/10.1038/d41586-022-00493-w) find
  53. Ecer, D. & Maciocci, G. Sciencebeam—using computer vision to extract pdf data. Elife Blog Post (2017). [Online; accessed 2023-Sep-8]. find
  54. Goodman, S. N., Berlin, J., Fletcher, S. W. & Fletcher, R. H. Manuscript quality before and after peer review and editing at annals of internal medicine. Annals internal medicine 121, 11–21 (1994). find
  55. Abid, A. et al. An online platform for interactive feedback in biomedical machine learning. Nat. Mach. Intell. 2, 86–88 (2020). find
  56. Collins, E., Augenstein, I. & Riedel, S. A supervised approach to extractive summarisation of scientific papers. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), 195–205 (2017). find
  57. Nuijten, M. B., Hartgerink, C. H., Van Assen, M. A., Epskamp, S. & Wicherts, J. M. The prevalence of statistical reporting errors in psychology (1985–2013). Behav. research methods 48, 1205–1226 (2016). find
  58. Zhou, Y., Beltagy, I., Bethard, S., Cotterell, R. & Chakraborty, T. ACL pubcheck. https://github.com/acl-org/aclpubcheck. ↗
  59. Zhang, J., Zhang, H., Deng, Z. & Roth, D. Investigating fairness disparities in peer review: A language model enhanced approach. arXiv preprint arXiv:2211.06398 (2022). find
  60. Hosseini, M. & Horbach, S. P. Fighting reviewer fatigue or amplifying bias? considerations and recommendations for use of chatgpt and other large language models in scholarly peer review. Res. Integr. Peer Rev. 8, 4 (2023). find
  61. Verharen, J. P. Chatgpt identifies gender disparities in scientific peer review. bioRxiv 2023–07 (2023). Parsed PDF Raw PDF Title Mind the Gap: Understanding Modality Gap... Abstract We present modality gap, an intriguing geometric… Introduction Multi-modal models map inputs from different data moda… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liang-2023-large-language-models-provide/