Impact of large language models on peer review opinions from a fine-grained perspective: Evidence from top conference proceedings in AI

1. Problem Statement

The academic peer review system faces increasing pressure due to rising submission volumes and a shortage of qualified reviewers, leading to issues like bias, inconsistent quality, and reviewer fatigue. The rapid emergence of Large Language Models (LLMs) has introduced a new variable: the potential for LLMs to assist or generate peer review reports.

While prior studies have established that LLMs are being used in peer review, critical gaps remain:

• It is unclear whether LLMs are altering the core evaluative functions of peer review (e.g., assessing originality vs. surface clarity).
• There is a lack of systematic examination regarding how LLMs affect the linguistic form, evaluative focus, and recommendation signals of review reports at a fine-grained level (word, sentence, and aspect levels).
• Major conferences (e.g., NeurIPS) lack explicit policies on LLM usage, creating an ambiguous environment where the impact of LLM assistance is difficult to quantify.

2. Methodology

The authors conducted a large-scale, fine-grained analysis of peer review reports from two top-tier AI conferences: ICLR (2017–2025) and NeurIPS (2016–2024, excluding 2020). The study utilized a multi-stage analytical framework:

A. Data Collection & Preprocessing

• Source: OpenReview dataset containing ~188,000 reviews and ~49,000 papers.
• Segmentation: Texts were segmented into words and sentences using NLTK.
• LLM Detection: To distinguish between human-written and LLM-assisted reviews, the authors employed:
  1. Maximum Likelihood Estimation (MLE): A model (adapted from Liang et al.) trained on expert-authored vs. LLM-generated reference texts to estimate the probability ($\alpha$) that a corpus contains LLM-modified content.
  2. Lexicon Filtering: A predefined dictionary of terms frequently used by LLMs (e.g., "meticulously," "seamlessly," "robustly") was used as a pre-filter. Reviews containing these terms were subjected to the MLE model.

B. Fine-Grained Analysis Dimensions

The study analyzed three distinct levels of granularity:

1. Word Level (Lexical Complexity): Measured using TAALES (e.g., COCA Academic Range, Bigram/Trigram frequencies) to assess vocabulary diversity and sophistication.
2. Sentence Level (Syntactic Complexity): Measured using TAASSC (e.g., adverbial clauses per clause, nominal subjects per clause, auxiliary verbs) to assess grammatical structure and logical flow.
3. Aspect Level (Content Focus): Using a pre-trained BERT-based sequence labeling model (optimized from Yuan et al.), reviews were annotated with eight specific evaluation aspects: Summary, Motivation, Originality, Soundness, Substance, Replicability, Meaningful Comparison, and Clarity. Sentiment polarity (positive/negative) was also annotated for each aspect.

C. Comparative Analysis

• Temporal Comparison: Pre-LLM (pre-2023) vs. Post-LLM (2023–2025).
• Group Comparison: LLM-assisted reviews vs. Non-LLM-assisted reviews.
• Confidence Stratification: Analysis was stratified by reviewer self-reported confidence scores (1–5) to determine if LLMs disproportionately affect low-confidence reviewers.
• Correlation Analysis: Spearman correlation coefficients were calculated between aspect mentions and both Overall Scores and Confidence Scores.

3. Key Contributions

1. Fine-Grained Linguistic Profiling: Unlike previous studies that analyzed reviews at the document level, this paper provides a granular breakdown of lexical, syntactic, and aspect-level changes in peer reviews following the advent of LLMs.
2. Detection of LLM-Assisted Patterns: The study identifies specific linguistic markers (e.g., increased nominal subjects, standardized sentence structures) and content shifts (e.g., reduced focus on originality) associated with LLM assistance.
3. Confidence-Dependent Impact: It reveals that LLMs have a more pronounced effect on reviewers with low confidence scores, helping them generate longer, more fluent reports but potentially shifting their focus away from deep critical reasoning.
4. Empirical Evidence on Evaluation Quality: The study challenges the assumption that LLMs drastically alter scoring outcomes, finding that while linguistic fluency increases, the correlation between aspect mentions and final scores remains weak.

4. Key Results

A. Linguistic Complexity and Structure

• Length: Post-LLM, review texts became longer and more fluent, particularly in ICLR 2025.
• Lexical Complexity: A slight decline in lexical sophistication was observed. LLMs tend to generate clear, readable text, avoiding overly complex terminology.
• Syntactic Patterns:
  • Nominal Subjects: Significant increase in the use of noun phrases as subjects (e.g., "The paper" vs. "I"), enhancing objectivity.
  • Auxiliary Verbs: A decrease in auxiliary verbs, suggesting more direct and concise sentence structures.
  • Standardization: Syntactic patterns became more uniform across different confidence levels, reducing individual stylistic variation.

B. Shift in Evaluative Focus (Aspect Level)

• Increased Emphasis: LLM-assisted reviews showed a marked increase in Summaries and Substance.
• Decreased Emphasis: There was a significant decline in mentions of Originality, Replicability, and Meaningful Comparison.
• Interpretation: LLMs appear to help reviewers articulate surface-level clarity and summarize content effectively but struggle to generate deep, critical insights regarding the novelty or reproducibility of the work.

C. Impact of Reviewer Confidence

• Low Confidence Reviewers (Score 1): Showed the most dramatic changes post-LLM. Their review lengths increased significantly, and they relied heavily on summaries. LLMs appear to act as a "scaffold" for these reviewers, allowing them to produce more informed-sounding reports.
• High Confidence Reviewers: Their reports became more concise and accurate, with less reliance on LLMs for structural generation.

D. Correlation with Scores and Confidence

• Scoring: The frequency of aspect mentions (even in LLM-assisted reviews) had a weak and statistically insignificant correlation with the overall scores assigned. This suggests LLMs are used for writing assistance rather than decision-making manipulation.
• Confidence: There is a weak negative correlation between Clarity mentions and reviewer confidence. When clarity is low, reviewers feel less confident even with LLM support.

5. Significance and Implications

• Theoretical: The study suggests that LLMs are reshaping the form of peer review (fluency, structure, length) more than its substance (scoring, deep critical evaluation). It highlights a potential trade-off: linguistic fluency is increasing, but evaluative depth (originality/replicability) is decreasing.
• Practical:
  • Policy: Conferences should establish guidelines that encourage LLM use for language polishing and summarization but restrict reliance for core evaluative judgments.
  • Tooling: Journals could provide official LLM-assisted tools to standardize quality and reduce bias, rather than leaving usage to individual discretion.
  • Future Research: The authors call for experimental studies to establish causality and broader analysis across non-AI fields (e.g., Nature, PLOS ONE) to verify if these trends are domain-specific.

Conclusion: The paper concludes that while LLMs are becoming integral to the peer review workflow, they currently function as efficiency tools that enhance readability and structure. However, they risk homogenizing review discourse and diminishing the focus on critical dimensions like originality, necessitating careful human oversight to maintain the rigor of academic evaluation.