CoChat

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

Open access

The DOI matches its DataCite record; 59 of the 60 references it cites were found in a registry or an index, and 1 are not checked yet; none of the 22 sources its claims cite is checked for support yet.

Abstract

Expert feedback lays the foundation of rigorous research. However, the rapid growth of scholarly production and intricate knowledge specialization challenge the conventional scientific feedback mechanisms. High-quality peer reviews are increasingly difficult to obtain. Researchers who are more junior or from under-resourced settings have especially hard times getting timely feedback. With the breakthrough of large language models (LLM) such as GPT-4, there is growing interest in using LLMs to generate scientific feedback on research manuscripts. However, the utility of LLM-generated feedback has not been systematically studied. To address this gap, we created an automated pipeline using GPT-4 to provide comments on the full PDFs of scientific papers. We evaluated the quality of GPT-4's feedback through two large-scale studies. We first quantitatively compared GPT-4's generated feedback with human peer reviewer feedback in 15 Nature family journals (3,096 papers in total) and the ICLR machine learning conference (1,709 papers). The overlap in the points raised by GPT-4 and by human reviewers (average overlap 30.85% for Nature journals, 39.23% for ICLR) is comparable to the overlap between two human reviewers (average overlap 28.58% for Nature journals, 35.25% for ICLR). The overlap between GPT-4 and human reviewers is larger for the weaker papers. We then conducted a prospective user study with 308 researchers from 110 US institutions in the field of AI and computational biology to understand how researchers perceive feedback generated by our GPT-4 system on their own papers. Overall, more than half (57.4%) of the users found GPT-4 generated feedback helpful/very helpful and 82.4% found it more beneficial than feedback from at least some human reviewers. While our findings show that LLM-generated feedback can help researchers, we also identify several limitations.

Checked 2026-11-19 · 1 reference not checked yet

Overview

CoChat Summary

TL;DR

This paper builds a GPT-4 pipeline that generates structured feedback on full scientific PDFs and evaluates it retrospectively against human peer review and prospectively with 308 researchers. It finds GPT-4 feedback overlaps with individual human reviewers at rates comparable to human–human overlap and is often perceived as helpful, but remains limited in specificity and in-depth methodological critique.

Key findings

  • Retrospective overlap was comparable to human–human overlap. In 15 Nature family journals (3,096 papers, 8,745 reviews), 57.55% of GPT-4 comments overlapped with at least one human reviewer, and 30.85% overlapped with an individual reviewer versus 28.58% for two human reviewers. In ICLR (1,709 papers, 6,506 reviews), 77.18% overlapped with at least one human reviewer, and 39.23% overlapped with an individual reviewer versus 35.25% for two human reviewers. Rejected ICLR papers showed the highest GPT-4–human overlap at 47.09% in Results (the abstract reports 43.80% for rejected ICLR papers; Results reports 43.80% as human–human overlap for rejected papers).
  • GPT-4 feedback was paper-specific rather than generic: shuffling reduced pairwise overlap from 30.85% to 0.43% on Nature family data and from 39.23% to 3.91% on ICLR. GPT-4 was more likely to echo comments raised by multiple reviewers: Nature family 11.39% for one reviewer, 20.67% for two, 31.67% for three or more; ICLR 15.39%, 26.21%, and 39.33%, respectively. Earlier comments were also more likely to overlap: in Nature family data, first-quarter comments had 21.23% overlap versus 16.74% for the second quarter.
  • GPT-4 emphasized different aspects than humans: it commented on implications of research 7.27 times more frequently than humans, was 10.69 times less likely to comment on novelty, humans were 6.71 times more likely to request additional ablation experiments, and GPT-4 was 2.19 times more likely to request experiments on more datasets.
  • In the prospective user study with 308 researchers from 110 US institutions, 57.4% found GPT-4 feedback helpful/very helpful (50.3% helpful, 7.1% very helpful), and 82.4% found it more beneficial than feedback from at least some human reviewers. 65.3% thought it at least to some extent offered perspectives overlooked or underemphasized by humans, and 50.5% were willing to reuse the system. More than 70% saw at least partial alignment with significant points/issues, with 35% seeing considerable or substantial alignment.

Methods

  • The authors developed an automated GPT-4 pipeline that parses a paper PDF with ScienceBeam, constructs a prompt from the title, abstract, figure/table captions, and main text, and generates structured feedback in four sections: significance and novelty, potential reasons for acceptance, potential reasons for rejection, and suggestions for improvement. Due to GPT-4’s 8,192-token input/output limit, the first 6,500 tokens were used, and feedback was generated in a single pass.
  • Retrospective evaluation used two datasets: 3,096 accepted papers and 8,745 human reviews from 15 Nature family journals, and 1,709 papers and 6,506 human reviews from ICLR 2022–2023, including accepted, rejected, and withdrawn papers. A two-stage comment-matching pipeline performed extractive summarization followed by semantic text matching; human verification yielded F1 = 96.8% for extraction and F1 = 82.4% for matching, with matches retained only at similarity ≥ 7.
  • Prospective evaluation recruited 308 researchers from 110 US institutions in AI and computational biology. Participants uploaded their own papers published after September 2021 and completed a 6-page survey; compensation was $20, and the study was approved by Stanford University’s Institutional Review Board.

Limitations

  • The paper states that results are based on one GPT-4 instantiation and one prompt, using zero-shot learning without fine-tuning, so findings should be interpreted as a lower bound. It also notes that evaluation covered only Nature family and ICLR papers at top English-language venues; the user study was limited to machine learning and computational biology and subject to self-selection; GPT-4 could not interpret visual data such as tables, graphs, and figures; and feedback was sometimes generic or vague, lacked in-depth critique of method design, and was biased toward certain aspects such as “add experiments on more datasets.”
  • The text makes evident at least one reporting inconsistency: the abstract states an average overlap of 43.80% for rejected ICLR papers, while Results reports 47.09% for GPT-4–human overlap on rejected ICLR papers and 43.80% for human–human overlap on rejected papers.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

DataCite · 10.48550/arxiv.2310.01783

Cites

59 of 60 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 0 Not valid · 14 Not checked · 125 Own results and interpretations

94 more claims

Create a free account to see all 139 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 20
  • Probably not a substantial claim. 25
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

13 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 60 references

Research works printed but not linked: 6

  1. Kuhn, T. S. The structure of scientifi revolutions. The Un Chic. Press. 2, 90 (1962).
  2. Nature. How to Write a Report. https://www.nature.com/nature/for-referees/how-to-write-a-report. Accessed: 21 September 2023.
  3. Rogers, A. & Augenstein, I. How to review for ACL Rolling Review. https://aclrollingreview.org/reviewertutorial (2021).
  4. Association for Computational Linguistics. ACL-IJCNLP 2021 Instructions for Reviewers. https://2021.aclweb. org/blog/instructions-for-reviewers/ (2021).
  5. International Conference on Machine Learning. ICML 2023 Reviewer Tutorial. https://icml.cc/Conferences/ 2023/ReviewerTutorial (2023).
  6. Zhou, Y., Beltagy, I., Bethard, S., Cotterell, R. & Chakraborty, T. ACL pubcheck. https://github.com/acl-org/ aclpubcheck.

Cited by

See all 633

Metadata

Authors
Liang, W., Zhang, Y., Cao, H., Wang, B., Ding, D., Yang, X., Vodrahalli, K., He, S., Smith, D., Yin, Y., McFarland, D., & Zou, J.
Affiliations
  • Stanford University (US)
  • Northwestern University (US)
  • Cornell University (US)
Venue
arXiv
Publisher
arXiv
Published
2023
Volume
1
Issue
8
Pages
AIoa2400196-AIoa2400196
ISSN
2836-9386 (ISSN-L)
OpenAlex
W4400734098
Cited by
633
References
33
Citations by year
2026: 310 · 2025: 205 · 2024: 108 · 2023: 9
FWCI
212.41
Type
journal article
Language
en
Open access
Locations
  • NEJM AI (publishedVersion)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Science Foundation (1763191, CAREER)
  • National Institutes of Health (1942926)
  • National Institute on Aging (P30AG059307)
  • National Institute of Mental Health (U01MH098953)
Topics
  • Social Sciences › Library and Information Sciences › Academic Writing and Publishing
  • Social Sciences › Library and Information Sciences › Publishing and Scholarly Communication
  • Social Sciences › Library and Information Sciences › Academic Publishing and Open Access
Keywords
large language models, LLM-generated feedback, peer review, research paper evaluation, GPT-4, manuscript preparation, scholarly communication, artificial intelligence research

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-16, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-16 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Can large language models provide useful feedback on research papers? A large-scale empirical analysis. https://app.cochat.ai/papers/10.48550/arxiv.2310.01783
The paper
Liang, W., Zhang, Y., Cao, H., Wang, B., Ding, D., Yang, X., Vodrahalli, K., He, S., Smith, D., Yin, Y., McFarland, D., & Zou, J. (2023). Can large language models provide useful feedback on research papers? A large-scale empirical analysis. arXiv. https://doi.org/10.48550/arxiv.2310.01783

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.