DiscoveryRSI: 27B model hits 6.3% on post-cutoff scientific hypotheses, beating GPT-6
omarsar0 · x · 2026-10-20
A new paper introduces DiscoveryRSI, applying recursive self-improvement to scientific hypothesis generation across four life-science domains. Its 27B model produced hypotheses matching post-cutoff published research 6.3% of the time, versus 2.5% for GPT-6 and 2.3% for Claude Opus 5. The system grows hypothesis trees via multi-agent evolutionary search, converts investigations into executable tasks, assigns credit after each run, and replays the research harness to mine productive strategies. Research quality rose from 4.86 to 5.39 with evolve/repair actions growing to 91.7%. Code and models are open-sourced.
More from AGI Musings
- Stardock CEO: local AI will dominate most AI work in 18 months, datacenters pointless — draginol · 2026-10-20
- Fields Medalist John Pardon: "The AI understands. 100%" — it now catches subtle errors in his notes — thesaraharminta · 2026-10-20
- Aaron Levie says agent swarms will unlock compute-bound work, but humans can't review them — dosco · 2026-10-20
- Sinofsky: AI struggles in companies because most workers can't flowchart their own jobs for agents — a16z · 2026-10-20
- The real new threat isn't a permanent underclass but the dissolution of the middle class — max_paperclips · 2026-10-20
- Joke about an alternate AI timeline where CEO Emmett Shear slowed everything down — wordgrammer · 2026-10-20