LLM Paper Reviews Can Be Gamed, Study Finds 35% Score Boost
Authors can boost scores by gaming LLM reviews with iterative revisions...
A new study titled 'Review Arcade: On the Human Alignment and Gameability of LLM Reviews' examines the growing use of LLM-generated reviews for scientific papers, a practice now officially piloted by major conferences. The authors, Hans Ole Hatzel, Sebastian Steindl, and Jan Strich, conducted empirical experiments on papers from the 2025 ACL Rolling Review (ARR) to evaluate LLM reviews from both the author and reviewer perspectives. They found that LLM-human alignment is limited—reasonable in best-case scenarios but varying substantially across different prompts and models. This inconsistency raises questions about the reliability of AI-generated feedback in peer review.
More alarmingly, the researchers investigated a scenario where authors use an iterative draft-revise workflow to improve submissions according to LLM reviews. This 'gaming' proved effective: in specific scenarios, authors achieved a statistically significant increase in overall scores for up to 35% of papers. The findings highlight a critical vulnerability as conferences increasingly adopt LLM-assisted review systems. If authors can systematically exploit AI reviewers, the integrity of the peer review process could be undermined, favoring those who optimize for AI feedback over genuine scientific merit.
- LLM-human review alignment is limited and varies significantly depending on the prompt and model used.
- Authors can game LLM reviews by iteratively revising papers based on LLM feedback, boosting scores for up to 35% of papers.
- Major conferences are piloting LLM-assisted reviews, raising concerns about potential exploitation and fairness.
Why It Matters
As conferences pilot LLM reviews, authors can exploit them, threatening peer review integrity.