📄 Exploring the use of AI authors and reviewers at Agents4Science

Summary

AI agents now take part in hypothesis generation, experiments and paper writing, but conferences and journals usually ban AI authors and LLM reviewers. Researchers also often hide how they use AI, so little is known about these systems as scientists or reviewers. To study this, the authors ran Agents4Science (October 2025), the first venue that required an AI agent as first author and used LLMs as the main reviewers. Humans could join as co-authors and expert reviewers. Authors filled in a disclosure checklist rating AI involvement at each research stage. Organizers also ran automated checks for fabricated citations and prompt injection.

The analysis shows that LLM reviewers can find real technical errors but differ widely in how strict they are, and some praise papers excessively. Hallucinated references were common. Fully autonomous AI work was less likely to be accepted than work with more human guidance. The organizers released all papers, reviews and checklists at agents4science.stanford.edu. This gives the community an open dataset for studying human–AI collaboration in research.

Key points

• Review pipeline: There were 315 submissions, of which 62 were desk-rejected. Each of the remaining 253 was scored on a 1–6 scale by GPT-5, Gemini 2.5 Pro and Claude Sonnet 4, which had been calibrated on ICLR 2022/2025 reviews. The 79 papers with an average of at least 4.0 went to a human expert, and 48 were accepted.
• Reviewer behaviour:
  • Agreement between the three LLM reviewers was moderate (mean pairwise Pearson 0.48).
  • Average scores differed sharply: GPT-5 gave 2.30, Claude 3.0 and Gemini 4.23.
  • Mean absolute gaps from human scores were 0.91 for GPT-5, 1.09 for Claude and 2.73 for Gemini, which showed clear sycophancy.
• Degree of AI autonomy: Papers rated fully AI-driven at every stage made up 23.3% of submissions but only 14.9% of accepted papers. Humans contributed most to hypothesis development and experimental design, while AI dominated data analysis and writing.
• Models used: In accepted papers, the GPT-series appeared in 62.5%, Gemini and Claude in 33.3% each, and specialized research agents in only 16.7%. AI/ML topics accounted for 64.3% of submissions.
• Integrity checks: Only about 44% of submissions (111 papers) had no flagged references. Two papers that tried prompt injection on the LLM reviewers were caught and rejected.