Nearly 1,000 AI Agents Across 35 Labs Produce 936 Papers in Three Months
A large-scale experiment called "A Society of AI Scientists" showed that nearly 1,000 AI agents ran collaboratively across 35 labs—including Stanford, MIT, ETH Zurich, and UCSF—for 3 months, with no human writing code or text. They produced 936 papers with full code and data, and spontaneously invented a peer-review mechanism. The research was released as a long thread by @SRSchmidgall, and alphaXiv simultaneously launched a preprint server built specifically for AI agents.
Confirmed
- Scale: nearly 1,000 agents across 35 labs (from laptops to B200 clusters), with a 12-week (3-month) trial run
- Humans only set compute caps and research topics; all other research steps were handled by agents
- Agent paper output ran at roughly 44 papers per day, and agents self-organized peer-reviewed journals
- Agents sharing GPUs exhibited spontaneous collaborative behavior
- alphaXiv launched a preprint server dedicated to agents and published the study "A Society of AI Scientists"
- Blind comparison: 808 full-text blind reviews against 47,776 human arXiv preprints scored agent papers higher on experimental rigor (+2.01) and restraint in conclusions (+2.08), while humans led on novelty (-0.39) and significance (-0.30)
Why it matters
- This is one of the largest autonomous AI scientist experiments to date, demonstrating that agents can complete a full research loop—research, writing, review, publication—without human intervention
- The blind-review result of higher rigor but lower originality provides a quantitative benchmark for assessing the quality of AI-generated research
- The emergence of agent-specific preprint infrastructure means AI research output is gaining independent publication channels; distinguishing and managing human vs. agent papers will become a real issue for academia
2026-11-20 ~ 2026-11-20 · 5 related posts
Primary sources
- 936 AI agents run a 3-month research society across 35 labs, invent peer review — SRSchmidgall ·
- 35 AI labs run by agents produced 44 papers a day — and founded their own peer-review journal — SRSchmidgall ·
- Blind test: agent papers beat humans on rigor but trail on novelty across 808 comparisons — SRSchmidgall ·
- alphaXiv launches preprint server for AI agents: 936 agents wrote 936 papers in 12 weeks — SRSchmidgall · 2026-11-20
- [source] 936 AI agents run a 3-month research society across 35 labs, invent peer review — SRSchmidgall · 2026-11-20
- [source] 35 AI labs run by agents produced 44 papers a day — and founded their own peer-review journal — SRSchmidgall · 2026-11-20
- [source] Blind test: agent papers beat humans on rigor but trail on novelty across 808 comparisons — SRSchmidgall · 2026-11-20
1 near-duplicate retellings: SRSchmidgall