DiscoveryRSI paper: 27B model's hypotheses hit 6.3%, beating GPT-6 and Claude Opus 5
Afinetheorem · x · 2026-10-19
Ling Yang's team (Princeton) introduces DiscoveryRSI (arXiv:2610.23839), a framework that recursively learns to generate valuable, testable scientific hypotheses.
- How it works: builds hypothesis trees via multi-agent evolutionary search, compiles selected investigations into executable Harbor tasks, and learns through three mechanisms — hypothesis credit optimization, harness replay to surface productive strategies and recurring failures, and rediscovery to transfer improvements to new investigations.
- Operational definition: a valuable hypothesis is evidence-grounded, targets consequential uncertainties, and includes feasible experiments, explicit decision criteria, and alternative explanations.
- New benchmark: DFM-Bench evaluates hypothesis quality at the levels of hypotheses, revisions, development paths, and research runs.
- Results: 6.3% of DiscoveryRSI-27B's hypotheses match later published work, vs 2.5% for GPT-6 and 2.3% for Claude Opus 5, across four life-science domains.
More from Research
- MIST Framework Separates Execution from Memory to Diagnose Video World Model Failures — hkuhk · 2026-10-19
- Dusting the Phantom: Gaussian Allocation by Spatiotemporal Destination Recovers Fast-Moving Detail — Minhyeok Roh · 2026-10-19
- EvoHeir paper: fresh evaluation cuts harmful inherited updates in self-evolving LLM loops by 68–83% — MBZUAI · 2026-10-19
- ReaDy Uses Recurrent States as Dynamic Context for Speculative Decoding in Hybrid Linear Attention LLMs — Bohan Sun · 2026-10-19
- PrefEval: preference-following accuracy falls below 10% at just 10 turns for most LLMs — math_rachel · 2026-10-19
- ReasoningTrap: diagnosing reasoning rigidity with doctored variants of famous puzzles — math_rachel · 2026-10-19