LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing

Research questions

• The study asks how distinguishable LLM-generated reviews are from human reviews and whether LLMs can identify deficient segments by jointly reasoning over submissions, reviews, and rebuttals.Its scope is NLP paper (meta-)reviewing, where rising submissions make it difficult to recruit enough reviewers.Source: §1 · p. 2

ReviewCritique dataset

• ReviewCritique pairs original NLP paper submissions with human-written and LLM-generated reviews annotated sentence by sentence for deficiency and professionalism, with explanations.Annotations were made by NLP experts, most with Ph.D. degrees or area-chair experience.Source: §1 · p. 2

LLMs as Reviewers

• LLMs produce more deficient review segments than humans and often deliver paper-unspecific, low-diversity reviews lacking constructive feedback.The study assesses review aspects including summaries, strengths, weaknesses, and writing, and proposes a metric for diversity across papers.Source: §1 · p. 2

LLMs as Metareviewers

• Even top-tier LLMs struggle to match human experts in assessing individual reviews, including identifying deficient segments and explaining their judgments.The benchmark covers both closed-source and open-source LLMs and frames meta-reviewing as knowledge- and reasoning-intensive rather than text summarization.Source: §1 · p. 2

Contributions

• The work contributes the dataset, the first sentence-level quantitative comparison of human and LLM reviews, and the first analysis of LLMs as reviewers and meta-reviewers.It presents the dataset as a resource for AI-assisted peer-review research and LLM benchmarking.Source: §1 · p. 2

2 Related Work

• Research on AI for reviews includes predicting scores, generating reviews and meta-reviews, and using NLP methods to evaluate human-review quality.Examples include score prediction (Wang et al., 2020), review generation (Gao et al., 2024; Wang et al., 2020; Liu and Shah, 2023), and meta-review generation (Li et al., 2023; Lin et al., 2023b).Source: §2 · p. 2
• PeerRead (Kang et al., 2018), MOPRD (Lin et al., 2023b), and NLPeer (Dycke et al., 2023) contain papers and reviews without expert annotations, while other datasets target specific review aspects.Peer Review Analyze annotates paper-section correspondence, aspect, functionality, and significance; ReviewCritique is presented as the first dataset to benchmark LLMs as responsible meta-reviewers.Source: §2 · pp. 2–3
• Prior studies examine LLM review deficiencies, while this work compares human-written and LLM-generated reviews at the sentence level and proposes a metric for review diversity.The fine-grained analysis is intended to identify specific areas where LLMs excel or struggle in generating high-quality reviews.Source: §2 · p. 3

3.1 Paper Submission & Review Collection

• The dataset includes 100 NLP submissions from ICLR and NeurIPS (2020–2023), with 3–5 reviews each, meta-reviews, rebuttals, and latest pre-deadline versions.Papers had publicly accessible human reviews and were selected for an equal accepted/rejected split; suspected AI-written human reviews triggered removal of the paper and its metadata.Source: §3.1 · p. 3
• For direct comparison, GPT-4, Gemini-1.5, and Claude Opus each generated three reviews for a balanced subset of 20 papers using ICLR review guidelines and example reviews.The 20-paper size reflected the time demands of annotation and was intended to allow an acceptable statistical comparison.Source: §3.1 · p. 3

3.2 Data Annotation

• Deficient segments include factual errors, unsupported criticism, unconstructive feedback, and overly subjective, emotional, or offensive judgments.The authors do not rely solely on rebuttals because they can be inaccurate, incomplete, or omit deficiencies outside the weaknesses portion of a review.Source: §3.2 · p. 3
• Forty NLP annotators with substantial publication and reviewing experience assessed reviews sentence by sentence after reading the original submission and related reviews, meta-reviews, and rebuttals.Two annotators independently labeled each paper’s reviews, and a senior area-chair expert resolved disagreements; separate annotators assessed LLM reviews without being told their origin, though complete unawareness was not guaranteed.Source: §3.2 · pp. 3–4

3.3 Data Statistics

• LLM reviews had more deficient instances than human reviews (100% vs. 71.57% at review level; 13.97% vs. 6.27% at segment level).For accepted submissions, LLM reviews had more deficient segments than for rejected submissions, unlike human reviews; the authors suggest human reviewers tailor weaknesses to perceived paper quality, while LLM criticisms are superficial and non-specific.Source: Table 1 · p. 4, §3.3 · pp. 4–5

3.4 Novelty of ReviewCritique

• ReviewCritique enables responsible meta-reviewer benchmarking through sentence-level expert labels and granular comparison of human- and LLM-written reviews.Annotators read the original submission, meta-reviews, reviews, and rebuttals before labeling, unlike prior datasets described in the comparison.Source: §3.4 · p. 5, Table 2 · p. 5

4.1.1 Error type analysis for deficiency

• Human review deficiencies often involve paper misunderstanding and inexpert critique, whereas LLM reviews more often suggest out-of-scope work and make fewer inexpert statements.LLM reviews also lack Missing Reference, Invalid Reference, and Concurrent Work errors because they do not cite specific works; their superficial, paper-unspecific style may help avoid expert-level mistakes.Source: §4.1.1 · pp. 5–6

4.1.2 Fine-grained review analysis

• LLM summaries have fewer inaccurate-summary errors, but their strengths often rephrase author claims rather than critically verify them.Inaccurate summaries make up 1.35% of LLM deficient segments (0.19% of all LLM segments), versus 5.75% of human deficient segments (0.36% of all human segments); 53.2% of LLM Strengths segments are rephrased from submissions. Human reviewers questioned 34.5% of LLM-generated Strengths segments for accepted papers and 51.9% for rejected papers.Source: §4.1.2 · p. 6

4.1.3 Review Diversity

• Human reviews have the highest ITF-IDF specificity score (6.04), followed by Claude Opus (4.09), Gemini (3.82), and GPT-4 (3.34).LLMs show a sharp diversity drop in the Clarity section, consistent with their tendency to praise all papers’ writing; GPT-4 has the lowest scores across review sections, suggesting more repetition.Source: Figure 1 · p. 7, §4.1.3 · p. 7
• Inter-LLM BERTScores range from 70.33 to 71.17, versus 59.15 between human reviewers, indicating less diversity among LLM reviews.The similar, high scores across LLM pairs suggest that using multiple LLMs may not substantially increase diversity of review perspectives and insights.Source: Figure 2 · p. 8, §4.1.3 · p. 8

4.2 LLMs as Metareviewers

• Closed-source models generally outperform open-source models on deficient-segment F1, but recall exceeds precision and overall F1 remains low.Claude Opus achieves the highest F1; GPT-4 and Gemini 1.5 perform slightly worse. Higher recall than precision across models and prompting strategies indicates a tendency to incorrectly flag segments as deficient.Source: §4.2 · p. 8, Table 4 · p. 9
• LLMs’ explanations for correctly identified deficient segments score relatively low overall, with Claude Opus strongest and open-source models generally weaker.GPT-4 and Gemini 1.5 show similar performance to Claude Opus, but all models struggle to articulate their reasoning in line with expert explanations.Source: §4.2 · p. 8, Table 5 · p. 9
• LLMs particularly struggle to detect Inaccurate Summary, Writing, Superficial Review, Experiment, Contradiction, and Unstated Statement deficiencies.These types fall below the models’ respective average recall: GPT-4 47.68%, Claude Opus 42.12%, and Gemini 1.5 50.37%; the difficulties align with their reviewer weaknesses, including superficial feedback and hallucinated or contradictory claims.Source: §4.2 · p. 9

• ReviewCritique contains human-written and LLM-generated reviews with detailed deficiency annotations and explanations.Source: §5 · p. 9
• LLMs can generate reviews, but often produce deficient, paper-unspecific segments that lack diversity and constructive feedback.Source: §5 · p. 9
• Even state-of-the-art LLMs struggle to assess review deficiencies effectively.Source: §5 · p. 9

Limitations

• ReviewCritique evaluates submission text but excludes figures, tables, and other visual elements.Including these components could provide a more comprehensive assessment of LLMs in peer review.Source: Limitations · p. 9
• The dataset is limited to NLP papers, so performance in other research areas and domain-specific challenges remain unassessed.Expanding the dataset to other domains could help assess the generalizability of the findings.Source: Limitations · p. 9
• The study assesses reviews based on original submissions before rebuttal, not author rebuttals or post-rebuttal reviews.Including these later turns could provide a more comprehensive understanding of LLMs across the peer review process.Source: Limitations · p. 9

Ethical Considerations

• LLMs are currently insufficient to replace human reviewers, especially where expert judgment and nuanced understanding are required.The study does not advocate using LLMs for paper (meta-)reviewing.Source: Ethical Considerations · p. 10
• Potential ethical risks include bias, lack of accountability, and undermining the integrity of scientific evaluations.Source: Ethical Considerations · p. 10

A.1 BERTScore

• For evaluating explanations of Deficient segments, the authors use microsoft/deberta-xlarge-mnli to compute BERTScore.This base model is used for evaluations reported in Tables 5, 10, and 11.Source: §A.1 · pp. 11–12
• For inter-LLM review complementarity, the authors use facebook/bart-large-mnli for BERTScore because its 1024-token context accommodates full reviews exceeding DeBERTa’s 512-token limit.DeBERTa is the officially suggested BERTScore base model; BART is selected for this experiment because some full reviews exceed DeBERTa’s input limit.Source: §A.1 · p. 12

A.2 Similarity Score in ITF-IDF

• ITF-IDF similarity is calculated with SentenceBERT using the fast, good-quality all-MiniLM-L6-v2 pretrained model, although any sentence similarity model can be used.The method cites Reimers and Gurevych (2019) for SentenceBERT.Source: §A.2 · p. 12

A.3 LLM Inference Details

• The closed-source models tested are GPT-4 (gpt-4-turbo), Gemini 1.5 (gemini-1.5-flash-latest), and Claude 3 (claude-3-opus-20240229).Source: §A.3 · p. 12
• The open-source models tested are Llama3-8B, Llama3-70B, and Qwen2-72B; GPT-4, Claude 3, Gemini 1.5, and Qwen2-72B receive the full prompt with instructions, paper title, body, and review.The prompt formats are shown in Tables 12 and 13.Source: §A.3 · p. 12
• Because Llama3-8B and Llama3-70B support at most 8k tokens, their prompts truncate the paper body while retaining the instruction, title, and review.Those retained prompt components are considered crucial for evaluation.Source: §A.3 · p. 12

B Influence of Different Thresholds in ITF-IDF

• Across different t values, ITF-IDF’s performance rank on full reviews remains the same.Table 6 reports the comparison across thresholds.Source: Table 6 · p. 12, B Influence of Different Thresholds in… · p. 12

C Error Types Detected by LLMs

• Table 7 reports the number and percentage of segments each LLM detects for each error type.Source: C Error Types Detected by LLMs · p. 12

E Explanation Score Across Different Prompts

• The explanation evaluation reports ROUGE and BERTScore for each LLM and prompt combination across Tables 10 and 11.Source: E Explanation Score Across Different… · p. 12, Table 10 · p. 16, Table 11 · p. 16

F Prompt Templates

• Tables 12 and 13 provide prompts for identifying deficient segments, while Table 14 provides the prompt for generating LLM reviews.Source: F Prompt Templates · p. 12, Table 12 · p. 17, Table 13 · p. 18
• Table 7 compares GPT-4, Claude, and Gemini on identifying deficient segments and marks error types with significantly lower percentages than the models’ average recall.Source: Table 7 · p. 13

D Deficient Segment Error Types

• Table 8 reports the percentage distribution of error types among deficient segments in human-written and LLM-generated reviews.Source: D Deficient Segment Error Types · p. 12, Table 8 · p. 14
• Table 9 lists the error types used to classify deficient review segments, each with an explanation defined by the annotation team.Source: D Deficient Segment Error Types · p. 12, Table 9 · p. 15

E Explanation Score Across Different Prompts

• Table 10 evaluates explanations for correctly identified deficient segments using the Labeling-All and Select-Deficient prompt methods.Source: Table 10 · p. 16
• Table 11 evaluates explanations using an ensemble of two prompt results, with final scores averaged across explanations generated by the prompts.Source: Table 11 · p. 16