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

A new survey warns that large language models being explored to ease the burden on scientific peer review are vulnerable to strategic manipulation, even as they show an ability to generate fluent critiques and approximate human reviewer scores [1][2]. The paper, submitted to arXiv on 23 Jun 2026, provides a systems-level analysis of LLM-based scientific peer review, focusing on two core evaluative functions: critique generation and score prediction [1][2]. It synthesizes findings across existing benchmarks and presents a taxonomy of modeling approaches, including prompt-based, supervised, retrieval-augmented, and alignment-optimized methods [2]. The authors note that while LLMs can produce fluent feedback, their reliability, robustness, and security as decision-support systems remain insufficiently understood [1][2]. Beyond performance metrics, the survey identifies several emerging robustness risks. These include prompt injection, data poisoning, retrieval vulnerabilities, and reward hacking, all of which expose automated review pipelines to potential strategic manipulation [1][2]. The study also highlights dataset constraints, evaluation shortcomings, and domain concentration biases that limit current assessment practices [2]. The push to automate peer review comes as the scientific community grapples with broader credibility challenges. The replication crisis, a term coined in the early 2010s, describes widespread failures to reproduce published scientific results, undermining trust in research across psychology, medicine, and other fields [3]. This crisis has given rise to metascience, a discipline that uses empirical methods to examine research practice itself [3]. Introducing LLMs into this high-stakes evaluation process adds a new layer of complexity, as the models themselves can be manipulated in ways that human reviewers cannot [2]. The survey reframes automated peer review as a high-stakes, multi-objective decision problem [2]. It outlines key open challenges in modeling subjective disagreement and cross-domain generalization, providing what the authors describe as a roadmap for developing robust, transparent, and trustworthy AI-assisted scientific evaluation systems [1][2]. The work arrives amid rapid expansion in the large language model sector, with models such as DeepSeek and Qwen being developed at reported costs far below earlier systems [8][10]. DeepSeek, a Chinese AI company, claimed it trained its V3 model for US$6 million, compared to the US$100 million cost for OpenAI's GPT-4 in 2023 [8].