AI-Driven Review Systems: Evaluating LLMs in Scalable and Bias-Aware Academic Reviews

• LLM reviewing remains challenging because models may hallucinate or persuade humans despite inaccuracies, making quality and appropriateness difficult to control.Source: §1 · p. 1
• Meta-prompting decomposes complex review tasks across expert LLM instances with tailored instructions and can outperform conventional prompting across tasks.Source: §1 · p. 1
• AI-based reviews can provide early feedback, support quality control, enable trend analysis, reduce human bias, and help readers identify high-quality papers.Source: §1 · pp. 1–2
• Scaling review is increasingly important as the volume of academic papers grows, including a 12.3% increase in arXiv submissions from 2022 to 2023.Source: §1 · p. 2
• The paper presents multiple LLM-based reviewing systems and supporting materials, including prompts, review questions, scores, evaluation results, and code.Source: §1 · p. 2

OpenReviewer

• OpenReviewer adapts reviews to venue and paper content while considering text, figures, ethics, and bias to target human-level quality and community-norm adherence.Source: OpenReviewer · p. 2
• OpenReviewer lets users upload papers, receive immediate review feedback and revision instructions, provide feedback, and submit revised versions.Source: OpenReviewer · p. 2, Figure 1 · p. 3

Papers with Reviews

• Papers with Reviews collects roughly 500 arXiv papers daily and 1,000 open-access Nature papers monthly, then reviews and evaluates them at scale.Source: Papers with Reviews · p. 3, Figure 2 · p. 3
• Papers with Reviews publishes LLM-generated reviews alongside human reviews to support research navigation, author improvement, and broader access to evaluation.Source: Papers with Reviews · p. 3

Reviewer Arena

• Reviewer Arena evaluates human and LLM reviews through direct anonymous pairwise comparisons, using preferences to compute win rates, reviewer scores, and rankings.Source: Reviewer Arena · pp. 3–4, Figure 3 · p. 4

Anonymous Human and LLM Evaluation

• Five expert evaluators compare anonymous reviews from a human, three closed LLMs, an open LLM, and GPT-4 Turbo.Each of 150 papers receives two randomly assigned reviews from the five reviewers.Source: Anonymous Human and LLM Evaluation · p. 4
• Reviewer quality is quantified from pairwise outcomes using win matrices, Bradley-Terry coefficients, and logistic regression.The win matrix records each competitor’s wins divided by total matches against another competitor.Source: Anonymous Human and LLM Evaluation · p. 4
• The Bradley-Terry model estimates relative reviewer strength with a logistic pairwise-win probability and ranks reviewers by descending fitted coefficients.Coefficients are estimated by minimizing binary cross-entropy subject to ξ1 = 0.Source: Anonymous Human and LLM Evaluation · pp. 4–5

Automatic LLM Prediction of Human Preference

• LLM-based autoevaluation uses synthetic judgments to increase effective sample efficiency while reducing reliance on human-labeled validation data.Prediction-powered inference estimates and corrects synthetic-data bias to improve evaluation precision without sacrificing accuracy.Source: Anonymous Human and LLM Evaluation · p. 5
• The study fine-tunes an LLM to predict which of two competing LLM responses humans prefer, treating the model as an adjudicator of alignment.The approach targets head-to-head preference prediction for review-quality evaluation.Source: Automatic LLM Prediction of Human… · p. 5
• Automatically injecting paper errors and comparing original versus modified reviews maps which shortcomings LLMs detect reliably and which define their review limitations.The analysis compares review scores and content before and after modification.Source: Automatic Discovery of LLM Review… · p. 6, Figure 6 · p. 6

Review Question Design

• The system supports venue-specific, paper-type-specific, adaptive, and adaptively generated review questions.For adaptive selection, the LLM chooses the top 10 questions from a bank of 40 based on paper content.Source: Review Questions · p. 6

Review Comparison

• GPT-4 extracts structured summary points from human and LLM reviews, then compares human-human and human-LLM overlap using similarity scores.Function calling is used to obtain structured responses and reduce spurious text.Source: Automatic Summary Comparisons · p. 6, Figure 7 · p. 6

Editorial Process Simulation

• Meta-prompting simulates the conference editorial process through personas including program chairs, area chairs, reviewers, and authors.The simulated process includes assignments, reviewing, rebuttals, discussions, meta-reviewing, decisions, and notification.Source: Role Playing: Dialogue Between LLMs · pp. 6–7, Figure 8 · p. 7

User Feedback

• User feedback combines five quantitative review-quality questions with an open-ended question about OpenReviewer outputs.Feedback is collected from users of Papers with Reviews and summarized in Appendix C.Source: §5 · p. 7

Limitations

• Venue-dependent context documents make score distributions resemble human distributions and use the full score range, but require yearly updates.The documents include prior-year statistics, reviewer and area-chair guidelines, ethics and conduct codes, and the formal review form.Source: §5 · p. 7
• Human preference prediction is limited because different people may prefer different reviews, although correcting for behavioral bias partially mitigates this gap.The authors use a Kaggle human-preference dataset as common ground for prediction.Source: §5 · p. 7
• The paper’s capability and error-testing analysis identifies limits of the current application, especially for understanding how human reviewers should use it.The authors state that these limitations are essential for appropriate application use and plan to study how authors use and trust LLM reviews.Source: §5 · p. 7

• LLM-generated reviews align well with human reviewers in blind human and automatic GPT-4 evaluations.Source: §6 · p. 7
• The study evaluates both review-quality alignment and adherence to conference guidelines and scientific norms, using human reviews as a baseline.Source: §6 · p. 7
• When given previous editorial decisions, the LLM aligns well with human reviewers, but detects overclaiming more reliably than failures to follow expected norms.Source: §6 · p. 7
• The authors conclude that iterative design and large-scale empirical evaluation are essential for calibrating LLM applications in reviewing.Source: §6 · p. 7
• The systems process tens of thousands of papers and aim to improve review efficiency, literature accessibility, and identification of high-quality papers and emerging trends.Source: §6 · p. 7

• Conference submission counts are reported across CVPR, ICML, NeurIPS, ICLR, and AAAI.Source: Figure 9 · p. 10
• The appendices provide tables listing collected papers by venue and mapping Nature journal identifiers to journal names.Source: Table 3 · p. 10, Table 4 · p. 10

User Feedback on Reviews

• User feedback scores range from 5 to 7, indicating that most reviews were rated very good to exceptional.Source: C User Feedback on Reviews · p. 11, Figure 10 · p. 11
• Readers praised the reviews as detailed, well organized, comprehensive, ethical, personalized, and useful for identifying drawbacks and improving papers.Source: C User Feedback on Reviews · p. 11
• Feedback also requested more elaboration in sections such as correctness and better balance between suggested additions and concision.Source: C User Feedback on Reviews · p. 11

Review Scores

• P1, P2, and P3 produced recommendation scores above the human average, while P4 reduced the score to 4.61 and P5 brought it near the human level.Source: D Review Scores · p. 13, Figure 14 · p. 13
• With all related context, GPT-4’s P5 score distributions resembled human distributions for most dimensions, but confidence remained skewed higher.Source: D Review Scores · p. 13

Evaluating Reviews

• The evaluation used a randomized 5% sample and assessed human and P5-written reviews on a 0–5 scale, where zero indicates a content-free review.Source: E Evaluating Reviews · p. 14, Table 5 · p. 14
• The confusion analysis found one high-scored GPT-4 acceptance rejected by humans and four low-scored GPT-4 rejections accepted by humans.Source: E Evaluating Reviews · p. 14

• Human evaluators scored human-written reviews 4.80 and GPT-4 reviews with all related contexts 4.76 for explaining review scores.Source: F Blind Human Evaluation · p. 14
• Reviews written by GPT-4 with all related contexts received the highest author-improvement score when evaluated by humans, at 4.79.Source: F Blind Human Evaluation · p. 14
• Human-written reviews evaluated by GPT-4 received the lowest author-improvement score, 4.14, among the reported conditions.Source: F Blind Human Evaluation · p. 14
• GPT-4 evaluated human-written reviews as more content-specific than human evaluators rated GPT-4-written reviews, with scores of 4.97 and 4.68 respectively.Source: F Blind Human Evaluation · p. 14

• Human and LLM scores for original papers are close, indicating general alignment before shortcomings are introduced.Figure 15 compares original-paper human and LLM scores with scores after different error injections.Source: G Synthesis and Analysis for Mapping… · p. 15, Figure 15 · p. 15
• Most data points show positive score differences after error injection, indicating that the LLM usually detects errors and shortcomings.The heatmap color intensity represents the magnitude of score changes between original and modified papers.Source: G Synthesis and Analysis for Mapping… · p. 15, Figure 16 · p. 16
• Detection is uneven: technical errors are mostly associated with low score differences, while baseline-comparison and metric omissions vary substantially across papers.Some papers show broad detection across categories, whereas others show predominantly low differences.Source: G Synthesis and Analysis for Mapping… · p. 15
• Adversarial edits delete, insert, or modify paper text to create controlled shortcomings for comparing original and modified LLM reviews.Examples include deleting related work, baselines, or citations; modifying equations; and introducing overclaiming.Source: G Synthesis and Analysis for Mapping… · p. 15, Figure 15 · p. 15
• The LLM markedly penalizes overclaiming and detects different error categories through reduced review scores.The category differences are sorted by magnitude relative to reviews of original papers.Source: G Synthesis and Analysis for Mapping… · p. 15, Figure 17 · p. 16

H Examples of Error and Shortcomings and Adversarial Prompts

• Deleting related work, baseline comparisons, or citations creates modified papers whose example review scores fall from 7 to 5, 4, and 5, respectively.The examples use side-by-side original and modified papers.Source: H Examples of Error and Shortcomings… · p. 17, Figure 18 · p. 20, Figure 19 · p. 21
• Overclaiming edits reduce the example review score from 7 to 3, providing the largest stated penalty among the listed examples.Figures 23–26 show original and modified overclaiming examples.Source: H Examples of Error and Shortcomings… · p. 17, Figure 23 · p. 25, Figure 24 · p. 26
• The proof discussion identifies a wrong theorem inequality: the proof establishes P(ℓa = ℓb) ≤ P(ℓb = ℓoracle), not a lower bound.The text says the proof logic is consistent but the theorem statement has the wrong inequality.Source: H Examples of Error and Shortcomings… · p. 17
• Theoretical mistakes and technical errors are subdivided into five sub-types, while adversarial prompts introduce eight error and shortcoming categories.The classification and prompt examples are summarized in Tables 6 and 7.Source: H Examples of Error and Shortcomings… · p. 17, Table 6 · p. 18, Table 7 · p. 19
• Modifying equations by removing square roots and flipping an inequality produces a case where ChatGPT detects the flipped inequality in Theorem 3.The original and modified equations are shown in Figures 21 and 22.Source: H Examples of Error and Shortcomings… · p. 17, Figure 21 · p. 23, Figure 22 · p. 24

I Automatic Comparisons

• Human–LLM review summaries have higher overlap than human–human summaries: 3.67 versus 3.05 average overlap.The reported standard deviations are 1.58 for human–LLM overlap and 1.56 for human–human overlap.Source: I Automatic Comparisons · p. 29
• The average weighted Jaccard similarity is 0.214, with Human–LLM similarities higher than Human–Human values.Common overlaps include experimental validation, methodological clarity, and potential real-world applications.Source: I Automatic Comparisons · p. 29, Figure 27 · p. 29
• GPT-4 simulates program chair, senior area chair, area chair, and reviewer roles, reducing the editorial process from human months to machine minutes.The role structure and responsibilities are described in Tables 8 and 9.Source: J Editorial Review Process · p. 30, Figure 28 · p. 30, Table 8 · p. 31
• The review-question framework considers conference- or journal-specific fixed questions and paper-type-specific fixed questions alongside other selection approaches.The paper treats question selection as a design choice affecting thoroughness and relevance.Source: K Review Questions · p. 33, Table 10 · p. 33, Table 11 · p. 34
• With all documents, GPT-4 score distributions resemble human distributions for correctness, novelty, significance, and overall recommendation, but confidence remains skewed higher.The overall recommendation distributions have similar means and standard deviations.Source: L Comparison of Score Distributions of… · p. 36, Figure 30 · p. 36, Figure 31 · p. 36
• The paper lists preventive actions for ethical and transparent use of LLMs in peer review.These actions are presented in Table 13.Source: M Preventive Actions for Ethical and… · p. 37, Table 13 · p. 37

N Predicting Human Preferences: Implementation Details

• Three open-weight models—Gemma-2-9b-it, Llama-3.1-8b, and Mistral-Nemo-Instruct-2407—are quantized to 4 bits for training and inference.The pipeline also applies data augmentation, hyperparameter tuning, and bias correction.Source: Fine Tuning · p. 38
• Training explores learning rates from 2e-6 to 2e-4, multiple optimizers, weight decay from 0 to 0.1, dropout from 0.03 to 0.12, and freezing 0, 10, or 16 layers.Inference varies sequence lengths from 512 to 8192 and batch sizes of 2, 4, or 8 across 100 trials per model.Source: Fine Tuning · p. 38
• The system augments the competition dataset by paraphrasing prompts while preserving associated responses.Anthropic’s prompt-generation tool and Phi-3-Mini-4k-Instruct are used for paraphrasing.Source: Data Augmentation for Fine-Tuning · p. 38
• The system exposes review and reviewer-arena interfaces for online use.The interfaces are shown in Figures 32 and 33.Source: Figure 32 · p. 39, Figure 33 · p. 39
• Bradley–Terry implementation builds win and count matrices, optimizes negative log likelihood, estimates competitor coefficients, and ranks competitors by descending coefficients.The first coefficient is fixed as a constraint during estimation.Source: P Code of Key Functions in Reviewer… · pp. 40–41
• The authors retain human involvement because sole LLM evaluation risks bias, favoritism toward similar LLM outputs, misinformation, and hallucinations.The stated goal is to combine humans and LLMs while examining the spectrum between human evaluation and full automation.Source: Q Levels of Autonomy in Reviewing · p. 42