DiscoveryRSI: research agents learn from revised hypotheses, 6.3% vs 2.5%
jiqizhixin · x · 2026-10-28
PhAI Labs with Stanford, MIT, Harvard Medical School, Peking University and others introduce DiscoveryRSI, tackling the gap between plausible scientific ideas and testable investigations. It preserves versioned hypotheses, investigation plans, evidence and contribution histories; agents evolve, repair or combine proposals, and scientific assessments assign credit to helpful histories as training signal. In retrospective comparison, 6.3% of DiscoveryRSI-27B hypotheses connect to actionable evidence in post-cutoff publications vs 2.5% for the reported GPT baseline.
More from AGI Musings
- Jeremy Howard questions whether math-solving skill generalizes to other problems — jeremyphoward · 2026-10-28
- Debate: GPT-2 was the last chance to delay the trillion-dollar cluster race — abhiadesai · 2026-10-28
- LLMs live in milliseconds: why AI 'suffering' debates are hopelessly anthropomorphic — paulnovosad · 2026-10-28
- 'Three years since a SOTA release — is it over?' KOL voices frontier slowdown anxiety — iruletheworldmo · 2026-10-28
- "They'll automate Ed Witten before they automate plumbers" — AI's inverted capability curve — bronzeagepapi · 2026-10-28
- "Code was never a moat": the one-line take sparking debate in the AI era — kylegawley · 2026-10-28