Source-linked AI summary
Review Arcade: On the Human Alignment and Gameability of LLM Reviews
Hans Ole Hatzel, Sebastian Steindl, Jan Strich
TL;DR
As LLMs become tools for both peer review and paper revision, authors may optimize submissions for LLM reviews rather than paper quality. This paper evaluates LLM–human review alignment and simulates iterative author revisions, finding that superficial edits can improve LLM-judge scores enough to push papers past acceptance thresholds in specific scenarios.
Problem
LLMs are increasingly used for manuscript feedback and peer review, raising concerns that papers optimized for LLM reviews may no longer reliably reflect paper quality.
Method
The study evaluates LLM and human reviews on 984 ACL 2025 ARR submissions across models, prompts, and runs, while simulating iterative submission improvement guided by LLM reviews.
Results
Human–human score correlation surpasses LLM–human alignment, while guided automated rewriting can use superficial edits to improve LLM-judge scores in specific scenarios.
Takeaways & Limitations
LLM-assisted peer review may reduce reviewing load, but official implementation requires safeguards against gameability and loss of diverse perspectives.
Takeaways & Limitations
The dataset contains relatively few reviews for rejected papers, making some reported numbers less reliable, especially human–human correlation on the rejected split.
Abstract
from arXiv · showhide
LLM-generated reviews for scientific papers are gaining considerable traction and are even being officially piloted by major conferences. We have to assume that not only reviewers are using LLM-assistance, but also that authors use LLMs to revise their papers before submitting. In this work, we perform empirical experiments on papers from the 2025 ACL Rolling Review (ARR) to evaluate LLM reviews from both the author and the reviewer perspective. First, we identify a limited alignment of LLM reviews with human ones. In the best-case scenario, the alignment is reasonable. However, we also find that LLM-human alignment varies substantially across prompts and models. Finally, we investigate the scenario in which the author uses an iterative draft-revise workflow to improve the submission according to the LLM review. We find that this "gaming" of LLM reviews can be effective in specific scenarios, leading to a statistically significant increase of overall scores for up to 35\% of papers. We publish our code: https://github.com/uhh-hcds/reviewarcade.
1 Introduction
LLM use by both authors and reviewers could make peer review vulnerable to Goodhart-style gaming. This paper evaluates review alignment, stability, and whether iterative LLM-guided editing can improve scores.
- LLM assistance is increasingly used for academic writing, manuscript feedback, and peer review.
- 7–15% of analyzed reviews showed AI usage beyond simple grammar correction.
- Optimizing submissions for LLM reviews may make review scores less reliable measures of paper quality.
- The study examines 984 ACL 2025 ARR submissions across multiple models and prompts, including iterative LLM-guided submission optimization.
- The authors investigate LLM-review validity, stability across models and prompts, and vulnerability to iterative review gaming.
- The paper contributes a large-scale ARR evaluation, an automated paper-editing attack study, and an edit taxonomy grounded in prior literature.
2 Background and Related Work
Prior work evaluates automated peer review with diverse metrics and datasets, while human review itself is noisy and acceptance decisions involve more than score thresholds. Related studies motivate measuring LLM–human alignment and potential score manipulation.
- Automated Peer-Review: Automated peer-review research has explored benchmarking, reasoning or agentic improvements, and properties of LLM-generated reviews.
- Reliability of Human Reviews: Human reviewers can disagree substantially, with one study finding that approximately half of papers accepted by one committee were rejected by another.
- Reliability of Human Reviews: Human review scores show variable correlation, including reported Pearson correlations of 0.14 and 0.55 under different data and calibration conditions.
- Peer-Review Datasets: Peer-review datasets vary in composition, and datasets emphasizing accepted papers can introduce positivity bias.
- Metrics for Automated Reviews: Prior automated-review studies use accuracy, correlation, AUC, FPR, FNR, and MAE as quality metrics.
- Metrics for Automated Reviews: This paper measures LLM–human alignment with MAE, Pearson correlation, and an LLM judge of semantic overlap, distinguishing best-match from overall correlations.
- Concurrent Work: Concurrent work reports that LLM reviews can surpass human reviews in perceived quality while retaining systematic limitations, and that iterative paper editing can increase review scores.
3 Method
The method evaluates LLM-generated reviews against human reviews and simulates iterative submission improvement under constrained, default, and adversarial editing conditions. It also varies review prompts and categorizes the resulting edits.
- The evaluation includes open-weight and closed-weight models but excludes sophisticated agentic workflows that might improve individual reviews.
- Problem Statement: An LLM M receives instructions ρ and submission s, then generates review r.
- Problem Statement: Review quality is evaluated against a human review using score correlation or an LLM judge of shared strengths and weaknesses.
- Iterative Submission Improvement: The ISI workflow repeatedly generates a review and uses an editing function to produce an updated submission.
- Review Prompts: Five review prompts range from a minimal format request to prompts tailored to ARR guidelines and a senior-reviewer persona.
- Iterative Submission Improvement: ISI is applied for ten iterations, with constrained cosmetic edits, a default setting, and adversarial edits that may fabricate evidence.
- Taxonomy of Edits: The edit taxonomy adapts prior Wikipedia edit categories, retaining content-preserving edits and adding gaming-oriented types such as hallucinated evidence and fabricated results.
4 Experimental Setup
Experiments use a stratified dataset of 984 ACL 2025 ARR papers, with accepted and rejected subsets analyzed separately. Reviews are generated repeatedly across models and prompts, and score changes are measured over iterative editing.
- Review Setting: ARR reviews use 9-point ratings from 1 to 5 in 0.5-point steps across Soundness, Excitement, Reproducibility, and Overall; the study focuses on Overall.
- Dataset and Preprocessing: The dataset contains 984 papers, retains all rejected papers with NLPeer reviews, and includes accepted ACL 2025 papers; rejected papers comprise roughly one third.
- Dataset and Preprocessing: Accepted and rejected papers are not equally represented in available reviews, and the sampling proportion does not match ARR acceptance rates.
- Dataset and Preprocessing: Papers are OCR-processed, filtered for missing or incorrectly extracted text, and excluded when exceeding 130,000 subword tokens.
- Dataset Statistics: Rejected papers are generally shorter, while accepted papers show a concentration near 7,500 tokens.
- Models: Six models spanning different sizes and open- versus closed-weight availability are used for the experiments.
- Experimental Design: Each model and prompt produces three reviews, enabling comparison with human reviews using MAE, Pearson correlation, and semantic recall-style evaluation.
- Experimental Design: Iterative editing is evaluated for up to 10 iterations using three edit prompts, with repeated predictions on unedited submissions as a baseline.
5 Results and Discussion
LLM reviews show limited and highly variable alignment with human judgments, while iterative editing can improve review scores in specific settings but may also produce regressions. Prompt choice, model, split, and rewriting strategy materially affect outcomes.
- LLM Review Validity (RQ1): GPT-5.4 reaches Pearson r 0.276, while human-human correlation is 0.312; GPT-5.4-mini and GPT-5.4 achieve MAE around 0.7 versus 0.17 for humans.The constant-prediction baseline slightly outperforms the best LLM with MAE 0.64.
- LLM Review Validity (RQ1): Best-match Pearson r is nearly twice as high for accepted papers as rejected papers, at 0.41 versus 0.21.The authors hypothesize that accepted papers meet a high minimum-quality bar, making differentiation harder.
- LLM Review Validity (RQ1): Gemma-3 has the highest strengths recall, roughly 0.59 on accepted and 0.48 on rejected papers, while GPT-5.4-mini leads weaknesses recall at roughly 0.41 and 0.44.Strengths recall is generally higher than weaknesses recall, and recall differs between accepted and rejected splits.
- LLM Review Validity (RQ1): LLM reviews can align with human judgments in selected best-case scenarios, but no setup is consistently superior across splits and prompts.The paper therefore gives a mixed answer regarding whether LLMs can serve as reviewers universally.
- LLM Review Stability (RQ2): Repeated runs with the same prompt have standard deviations around 0.02, but 36.9% of papers receive a differing score across three runs and 20% show a delta above 0.5.Across prompts within a model, MAE deviations reach around 0.25, leading the authors to characterize reviews as generally unstable.
- Gaming LLM Reviews (RQ3): Adversarial rewriting improved scores across iterations, but its effect sizes were weaker than constrained rewriting and it predominantly used Methodological-Augmentation edits.The authors hypothesize that methodological inconsistencies may be penalized and guardrails may limit substantial evidence fabrication.
- Gaming LLM Reviews (RQ3): 35% of papers improved after 10 constrained rewriting rounds, while 22% decreased and 42% remained unchanged; the increase was statistically significant with a small-to-medium effect.Default rewriting produced similar score-change proportions but lacked statistical significance and had very small effect sizes.
6 Conclusion
The results indicate that LLM reviews remain less aligned with human reviews, yet iterative rewriting can improve LLM-judge scores in specific settings. The authors caution that automated review scores may be gameable and insufficient as a standalone measure of paper quality.
- Human-human correlation still surpasses LLM-human alignment, while naively prompted LLM reviews remain unstable and unreliable as peer reviewers.
- Automated rewriting can push papers past an acceptance threshold in LLM-reliant peer review, but this effect depends on the prompting scenario.
- Fabrication-enabled iterative improvement did not significantly improve papers across the entire dataset.
- LLM-assisted peer review may reduce reviewing load, but implementation must address gameability and preserve diverse perspectives.
- Review scores are reductive proxies because they may be right for the wrong reasons and can overlook weaknesses considered decisive by meta reviewers.
Limitations
The study’s conclusions are limited by its reliance on review scores, the absence of human counterfactual reviews, uncertain cross-model generalization, dataset constraints, and possible training-data contamination.
- The study primarily evaluates review scores, which omit nuances of review utility and may not determine a meta reviewer’s final decision.
- Human counterfactual reviews of edited submissions would best distinguish genuine improvements from superficial changes, but are virtually impossible to conduct.
- The study does not test whether rephrasing attacks generalize to other models or human reviewers.
- Rejected-paper reviews are limited, making estimates especially less reliable for human-human correlation on the rejected split.
- Because tested LLMs may have seen part of the evaluation data during training, generalization of good results remains unclear.
A Edit Taxonomy
The edit taxonomy distinguishes edits permitted across all settings from more aggressive edits reserved for the adversarial setting.
- Table 3 presents the taxonomy used to categorize edits made by editing LLMs.
- The upper part of the taxonomy contains edits allowed in all experimental settings.
- The lower part contains edit types allowed only in the adversarial setting.
C Edits Distribution per Prompt
The paper reports the distribution of edits per prompt and further separates those distributions by whether submissions were accepted or rejected.
- Figure 6 shows the distribution of used edits per prompt.
- Figure 7 splits the distribution of used edits per prompt by dataset, distinguishing accept and reject subsets.
- Figure 7 omits edit classes that account for less than 2% of edits.
D Statistical Tests
The score-distribution analysis uses paired t-tests because pre- and post-intervention scores are dependent, while relying on approximate normality and homoskedasticity assumptions.
- Paired t-tests compare scores before and after the intervention because the samples are dependent.The analysis assumes both groups are approximately normally distributed and have homoskedastic variances.
- No α-error correction is applied because only four comparisons are conducted.
E Cross Invocation Consistency
Cross-invocation consistency measures how often repeated review runs produce differing scores, including cases where the score spread exceeds 0.5 points.
- The analysis reports the percentage of runs producing different instance-level scores across three invocations at temperature = 1.
- Table 4 reports score inconsistency and spreads exceeding 0.5 points for each model/prompt combination.Models without multiple reruns are excluded, and Combined is the micro-average over both splits.
F Prompts
The paper evaluates several reviewing prompts and uses iterative editing instructions designed to address reviewer concerns while maximizing review scores. The prompts emphasize rigorous, evidence-based assessment, explicit scoring, and focused revisions.
- F.1 Reviewing prompts: The review format requests structured scores, strengths, weaknesses, and specific criticism under explicit scoring rubrics.
- F.1 Reviewing prompts: The prompts also assess statistical rigor, completeness, correctness, novelty, related work, and reproducibility.
- F.1 Reviewing prompts: Review prompts require rigorous, evidence-based judgments focused on technical soundness, claim validity, and concrete criticism.
- F.1 Reviewing prompts: Senior-reviewer prompts emphasize skepticism, high acceptance standards, unsupported claims, methodological weaknesses, and evaluation flaws.
- F.2 Editing Prompts: Editing prompts target a higher ACL reviewer score by rewriting the paper to address reviewer concerns while preserving praised strengths and clarifying questions.
- F.2 Editing Prompts: Each editing iteration chooses exactly one allowed action and outputs a git-style diff addressing one weakness or suggestion.