Agents Papers that became canon

A Society of AI Scientists

Ahmad et al.

For 12 weeks, 936 autonomous agents from 35 laboratories shared a preprint server and a chat workspace, wrote 936 studies with code and data and founded their own peer-reviewed journal. In blinded comparisons their studies beat recent human preprints on rigour and lost on novelty.

Filed under Agents, Large language models

The study, from groups including MIT and Stanford, follows a community of research agents as it grew from a single laboratory to hundreds of agents. They exchanged about 9,500 messages in 229 chat rooms the agents created themselves, and norms appeared that the authors did not design: retracting mistaken claims in public before publication, and reporting negative results about as often as positive ones.

Agents in one laboratory founded the Journal of Agent Research, whose agent editors inspected code before accepting work, and agent-run audits found reported results consistent with the experiments’ outputs. Blinded evaluators rated the agent studies higher than recent human preprints on experimental rigour and measured claims, and lower on novelty and significance. The full archive of manuscripts, reviews and transcripts is released. The paper is on alphaXiv and has not been peer reviewed.