acceptodds
Under review as a conference paper at ICLR 2027

LLM-REVal: Can We Trust LLM Reviewers Yet?

Abstract

The rapid advancement of large language models (LLMs) has encouraged their integration into academic workflows, potentially reshaping how research is conducted and reviewed. In this study, we examine how the joint use of LLMs in research generation and peer review may affect scholarly fairness, focusing on the risks of using LLMs as reviewers through simulation. Our simulation incorporates a research agent that generates and revises papers, alongside a review agent that evaluates the submissions. Based on the simulation results, we conduct human annotations and identify notable misalignment between LLM-based reviews and human judgments: (1) LLM reviewers assign more favorable scores to LLM-authored papers than to matched human-authored papers; (2) some limitation-oriented human-authored papers remain below the acceptance threshold even after multiple revisions. Additional analyses further characterize these disparities through paper-level linguistic profiles, the effect of LLM polishing, and limitation-oriented topics. These results highlight risks and equity concerns for human authors and scholarly evaluation if LLMs are deployed in peer review without adequate caution.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI in research, AI in review, systematic bias, fairness

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yAu4OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper examines whether LLMs can function as reliable reviewers in academic peer review. To explore this, the authors develop a multi-round simulation framework called LLM-REVal, which models the interaction between a Research Agent—responsible for generating both human-like and LLM-authored papers—and a Review Agent that evaluates those submissions. The simulation reproduces the full review process, including initial assessment, rebuttals, revisions, and resubmission cycles, with a focus on fairness and bias in LLM-based reviewing.

The results reveal that LLM reviewers systematically favor the writing style characteristic of LLM-generated papers, leading to inflated scores for LLM-authored work. In contrast, some human-written papers—particularly those that discuss risks, fairness, or critical perspectives on AI—consistently receive lower scores, even after multiple revisions.

Critical strengths
  • Novel Problem Formulation: Addresses a timely and underexplored issue—what happens when LLMs act as both researchers and reviewers, creating feedback loops in scientific workflows.
  • Comprehensive Simulation Framework: The paper develops a realistic multi-agent system encompassing literature search, paper creation, feedback, rebuttal, revision, and meta-review. This end-to-end simulation is technically impressive.
  • Clear Empirical Evidence of Bias: The study rigorously demonstrates two forms of bias in LLM reviewers:
    • Linguistic bias toward LLM-style writing.
    • Topic/framing bias against papers emphasizing risks or fairness.
Critical weaknesses

No critical weaknesses from my perspective. I reviewed this paper last year, and I am still supportive of it. I think the authors updated the paper following the guidance of other reviewers, so I believe we can consider accepting it.

Questions for authors

NA

Reviewer_rMi7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents a simulated process in which LLMs author papers (matched with human paper of the same topic and keywords), then an LLM reviewer pipeline engages in a review process (providing reviews, getting rebuttal, providing meta-review). This is done in successive revise-and-resubmit rounds as well. The paper finds that the LLM review process favors LLM-authored papers over human-authored papers, that the revise-and-resubmit process helps both for human-authored and LLM-authored papers, and that the LLM review process penalizes, relative to human preference, papers about limitation-oriented topics.

Critical strengths

This is a timely area of study, with a very interesting setup, testing not only LLM preferences of human- vs LLM-generated papers but also a revise-and-resubmit loop. Strong motivation and creative setup helps my overall assessment.

Several results are strong: that LLMs favor LLM-gen over human across several backbone models is very useful to see, and the LLM reviewer penalty to limitation-oriented topics is very directly relevant to a world that increasingly uses LLMs to review papers. The urgency of that last result, in particular, sways my overall assessment upward.

Critical weaknesses

I have concern about the validity of the way authors are using LLMs to generate results to be used in the paper. I understand that it might well be prohibitive to actually run the experiments, and I think that it is very good that the paper caps each predicted numerical improvement over baselines by the corresponding magnitude observed in the matched human-authored paper, but I have a hard time understanding how this matching process can be done in a meaningful way, as stated. The LLM-generated papers potentially have completely different ideas and experiment designs relative to the matched human papers, so it is unclear how the cap process keeps the LLM-generated papers from having much more impressive results (e.g. by proposing a much more moonshot-y experiment than the human paper, and showing success). I think this could be ameliorated by, at least, post hoc demonstrating some control over this matching beyond a simple cap. I might have phrased this as a question for the authors (how can they ensure some level of equivalence), but I really see this as a central issue informing how to interpret these results.

This carries over into a concern about how to interpret the revise-and-resubmit experiment. In this case, how is the cap meaningfully enforced? Both human papers and LLM papers are then revised by LLMs -- do they need to keep the same set of numerical results, or are they allowed to change them? This is a lesser concern but still bares on how to interpret the results.

Questions for authors

Everything asked in the weaknesses part.

Reviewer_ea2DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LLM-REVal simulates paper generation, review, and revision to study LLM reviewer behavior. Across three reviewer models, LLM-authored papers receive higher scores, revisions improve scores, and some limitation-oriented human papers remain poorly rated. Human checks and linguistic analyses examine these disparities.

Critical strengths
  • Studying generation and review together addresses a meaningful risk beyond evaluating isolated LLM reviews.
  • Similar scoring disparities appear across three reviewer models, supported by targeted human checks and content-preserving polishing experiments.
Critical weaknesses
  • Topic matching does not establish comparable scientific quality. Generated papers use different ideas and predicted results; capping numerical improvements does not remove these confounds.
  • Human validation covers only the 15 largest LLM-favoring gaps. Annotators also see papers without figures, references, or appendices, unlike the LLM reviewers, limiting broader conclusions.
  • Revision gains lack an unchanged-paper rescoring control. Selecting initially rejected papers introduces regression-to-the-mean concerns, so score increases cannot be fully attributed to revision quality.
Questions for authors
  • Does the authorship preference persist when scientific content is held fixed and humans assess both versions?
  • Does human–LLM disagreement persist on randomly selected, complete-paper pairs?
  • How much revision gain exceeds unchanged-paper rescoring? Please also specify the paper subset, reviewer backbone, and source of the pre-revision scores in Table 3, and explain their relationship to Table 2.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.