LLM-Based Scientific Peer Review: Methods, Benchmarks, and Reliability Challenges

Scientific publishing is facing a scalability crisis as the volume of submissions grows significantly faster than the available pool of qualified reviewers. This pressure has led to increased interest in using Large Language Models (LLMs) to automate two critical tasks: generating textual critiques and predicting numerical scores. While early computational approaches relied on feature-based prediction, modern LLMs are now being deployed to produce full-length reviews that mimic human-like reasoning and evaluation.

The authors categorize current LLM-based peer review systems into four primary paradigms:

• Prompt-based systems: These leverage the inherent instruction-following capabilities of general-purpose LLMs. While highly accessible, they are prone to sensitivity in prompt wording and often struggle with score calibration.
• Supervised fine-tuning: By training models on historical datasets of manuscripts and their corresponding expert reviews, these systems aim to better align with specific domain norms, though they risk inheriting the biases and noise present in human-written reviews.
• Retrieval-augmented generation (RAG): These systems incorporate external knowledge bases to ground critiques in factual evidence, potentially reducing hallucinations.
• Alignment-optimized approaches: These focus on refining model outputs to match human preferences, often using reinforcement learning to ensure that the generated critiques are both constructive and logically consistent with assigned scores.

Beyond performance metrics, the authors highlight that automated peer review is a high-stakes, multi-objective decision problem. Current systems face significant risks, including:

• Robustness and Security: Automated pipelines are susceptible to prompt injection, data poisoning, and reward hacking, where strategic actors might manipulate the system to favor specific submissions.
• Subjectivity and Disagreement: Peer review is inherently noisy; different human reviewers often disagree on the same manuscript. LLMs struggle to model this uncertainty, often producing overly confident or generic assessments that fail to capture the nuance of expert debate.
• Domain Generalization: Models trained on data from one field (e.g., computer science) may fail to adapt to the specific evaluative criteria or stylistic expectations of other scientific disciplines.