Johns Hopkins paper: LLMs make more scientific discoveries when training facts appear only once
rohanpaul_ai · x · 2026-11-06
- A new Johns Hopkins paper (arXiv:2611.03617) argues that the share of "monofacts" — facts seen exactly once during training — governs an LLM's scientific discovery capacity.
- Fine-tuning gpt-oss-20b and gemma-3-12b-it on materials science and regulatory genomics datasets, the authors varied monofact rates from 0% to 100%: discovery rates (exact recovery of held-out materials or DNA sequences) rose steadily in every setting, across sampling strategies and reasoning budgets, and the effect can't be explained by diversity alone.
- Training 3 or 5 epochs, which turns single sightings into repeats, erased the trend.
- Since discovery and hallucination both mean producing unseen content, more discoveries come with more hallucinations. Practical takeaway: keep rare one-off examples and avoid repeating data when building science models.
Related event: Study Links Share of Once-Seen Facts to LLM Scientific Discovery(2 posts)→
More from Research
- Agile Deliberation code released: Demo notebook walks through the full concept deliberation workflow — aigclink · 2026-11-06
- Google open-sources Agile Deliberation, an AI framework that turns fuzzy classification rules into explicit ones — aigclink · 2026-11-06
- Systematic review: GenAI fosters or hinders critical thinking depending on how it's used — _akpiper · 2026-11-06
- Developer Validates AME-2 on Humanoids: 70cm Foot-Only Climbing, 1.2m Gap Traversal in Sim — ChongZzZhang · 2026-11-06
- Citizen science labeling works better when small groups deliberate, not annotate alone — _akpiper · 2026-11-06
- Wayve releases Learned Safety strategy after a decade of end-to-end AV research — alexgkendall · 2026-11-06