How Closely Do LLM Reviews Align with Human Peer Review?

Technical Summary: Alignment of LLM Reviews with Human Peer Review

Problem Statement
Large language models (LLMs) are increasingly deployed to generate scientific reviews, yet existing evaluations often fail to determine whether different model providers align with both conference acceptance decisions and human reviewing priorities within a single controlled setting. A critical gap exists in understanding whether an automated reviewer's ability to distinguish accepted from rejected papers implies alignment with finer-grained human judgments, such as the distinction between oral and poster presentations, or whether it shares the same thematic priorities regarding paper weaknesses.

Methodology
The study conducted a controlled, cross-provider evaluation using 300 topic-matched submissions from ICLR 2026. The dataset comprised 100 oral, 100 poster, and 100 rejected papers. To ensure topical consistency, oral papers were matched with thematically similar poster and rejected papers using abstract embeddings (OpenAI text-embedding-3-small) and one-to-one similarity matching.

Three contemporary LLMs were evaluated: OpenAI gpt-5.4, Google gemini-3.1-pro-preview, and Anthropic claude-opus-4-6. To mitigate data leakage, models were selected based on training cutoffs predating the ICLR 2026 submission period, and external search was disabled. Each model reviewed every paper using identical instructions and a standardized rating scale (0–10) after decision information was removed from the PDFs.

The analysis focused on three dimensions:

1. Outcome Alignment: Whether LLMs assign systematically different ratings to accepted (oral/poster) versus rejected papers, and whether they distinguish between oral and poster categories.
2. Scale Usage: Comparing rating distributions and the specific usage of the recommendation scale across providers and against human baselines.
3. Thematic Agreement: Analyzing written weaknesses through embedding-based thematic clustering (UMAP dimensionality reduction followed by k-means clustering) to identify prioritized criticism themes.

Statistical significance was assessed using Mann–Whitney U tests for distribution differences and paired Wilcoxon signed-rank tests for direct comparisons between evaluators, with Holm correction applied for multiple comparisons.

Key Results

• Broad vs. Fine-Grained Discrimination: All three LLMs successfully distinguished accepted papers from rejected papers, assigning significantly higher ratings to the former. However, none of the models reproduced the distinction between oral and poster papers observed in human ratings. Human reviewers showed a clear median rating progression (Reject: 4.0 → Poster: 5.33 → Oral: 6.5), whereas LLMs failed to differentiate between poster and oral categories.
• Provider-Specific Scoring Patterns: Scoring behaviors varied significantly by provider. Google Gemini assigned systematically higher ratings across all categories, including assigning a score of 8 to 16 rejected papers. OpenAI and Claude were closer to human ratings for rejected and poster papers but were more critical (assigned lower ratings) than humans when evaluating oral papers.
• Scale Usage: LLMs utilized a narrower portion of the rating scale compared to humans. None of the models assigned the extreme endpoints (0 or 10), which were used by human reviewers.
• Thematic Divergence: While human and LLM reviews shared major weakness themes (e.g., experimental evidence, theoretical analysis), their emphasis differed. LLMs more frequently identified "insufficient baseline comparisons" as a primary weakness. In contrast, human reviewers more often raised concerns regarding "computational efficiency and resource requirements."

Contributions
This work provides three primary contributions:

1. A controlled cross-provider evaluation of LLM-generated peer-review scores across both broad acceptance outcomes and finer presentation categories.
2. Evidence that distinguishing accepted from rejected papers does not imply alignment with finer conference decisions or equivalent usage of the recommendation scale.
3. A comparison of thematic priorities, demonstrating that disagreement between humans and LLMs extends beyond numerical ratings to the specific weaknesses considered salient.

Significance and Claims
The paper concludes that while LLMs can serve as sources of supplementary critique—capable of identifying recurring weaknesses and providing feedback from multiple perspectives—their scores and priorities should not be treated as interchangeable with human peer-review judgments. The findings suggest that broad decision alignment does not guarantee agreement with finer human judgments. Consequently, future automated-review systems should be evaluated not merely by their fluency or ability to separate accepted from rejected papers, but also by their stability, scale usage, evidence grounding, and alignment with the nuanced structure of expert evaluation. The authors caution against relying on LLMs for final decision-making without acknowledging these systematic divergences.