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

LLMs show promise but lack proven reliability for peer review: While recent studies demonstrate that LLMs can generate fluent critiques and approximate reviewer scores, the paper emphasizes that "their reliability, robustness, and security as decision-support systems remain insufficiently understood"—a critical gap given peer review's importance in scientific gatekeeping.

Automated review systems are vulnerable to manipulation attacks: The research identifies specific security risks including prompt injection, data poisoning, retrieval vulnerabilities, and reward hacking that could allow bad actors to strategically manipulate automated review pipelines, raising concerns about the integrity of scientific evaluation if these systems are deployed without safeguards.

Current evaluation methods have significant blind spots: The paper identifies "dataset constraints, evaluation shortcomings, and domain concentration biases that limit current assessment practices," meaning existing benchmarks may not adequately test whether LLM reviewers work fairly across different scientific fields or handle edge cases that real peer review encounters.

Disagreement between reviewers remains a major unsolved problem: The survey highlights that modeling "subjective disagreement" is an open challenge—since human reviewers often disagree on paper quality, LLMs trained on limited consensus data may fail to capture legitimate scientific debate or produce artificially narrow evaluations.

LLM peer review requires a multi-objective approach, not just accuracy metrics: The paper reframes automated peer review as a "high-stakes, multi-objective decision problem," suggesting that optimizing solely for score prediction accuracy misses critical dimensions like transparency, fairness, and trustworthiness needed for responsible deployment in scientific publishing.