DiscoveryRSI recursively improves scientific hypothesis generation
Original title: DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
Automated scores rose from 4.86 to 5.39 after two self-improvement rounds, with larger gains in falsifiability and data resolvability.
The paper introduces DiscoveryRSI, which builds hypothesis trees with multi-agent evolutionary search, compiles investigations into executable Harbor tasks, and learns through credit optimization, harness replay, and rediscovery. DFM-Bench assesses hypotheses, revisions, paths, and runs across four life-science domains. Preliminary expert review rates DiscoveryRSI-27B above GPT-6 Luna on importance and evidence-bounded plausibility, and above GPT-6 and Claude Opus 5 on post-cutoff correspondence.
Unconfirmed: Human comparisons are described as preliminary assessments by the authors.
- Hypothesis quality rose from 4.86 to 5.39 after two rounds.
- DiscoveryRSI-27B matched later published research at a 6.3% rate.
- GPT-6 scored 2.5% and Claude Opus 5 scored 2.3% on that measure.
- Learning uses credit optimization, harness replay, and rediscovery.
From the original article
1 min read at arxiv.orgScientific discovery requires developing worthwhile hypotheses into practical investigations, yet learning from this evolving research process remains challenging. We introduce an operational definition of valuable and testable research hypotheses: evidence-grounded claims about consequential uncertainties, coupled with feasible investigations, explicit decision criteria, and alternative explanations.
By Jianyuan Zhong · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.