Reproduce “Pre-review to Peer review: Pitfalls of Automating Reviews using Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Akhil Pandey Akella, Harish Varma Siravuri, Shaurya Rohatgi
A multidisciplinary study evaluated frontier Large Language Models for scientific peer review, finding that while current models are unsuitable for autonomous review due to systematic overconfidence and inflated scores, their judgments correlate unexpectedly strongly with post-publication citation impact, suggesting a potential role in pre-screening. The work identifies crucial misalignments between LLM and human evaluation and biases against disruptive research.
Key insights
- Large Language Models consistently exhibit high overconfidence and a strong tendency towards positive score inflation, leading to inflated ratings for low-quality papers and a failure to express calibrated uncertainty.
- LLM-generated review scores paradoxically show stronger correlations with 5-year citation counts than human reviews but also a more pronounced negative bias against disruptive scientific work.
- While LLMs can generate reviews semantically similar to human ones, prompt engineering can only partially mitigate systematic scoring biases, particularly for identifying weaker submissions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.