DiscoveryRSI recursively learns to generate research hypotheses
Original title: DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
Automated scores rose from 4.86 to 5.39 after two self-improvement rounds across four life-science domains.
The paper introduces DiscoveryRSI, a multi-agent framework that grows hypothesis trees, compiles selected investigations into executable Harbor tasks, and learns from hypothesis credit, harness replay, and rediscovery. It also introduces DFM-Bench for scoring hypotheses, revisions, development paths, and research runs. Automated evaluation across four life-science domains shows hypothesis quality rising from 4.86 for the base model to 5.39 after two rounds of self-improvement, with larger gains in falsifiability and data resolvability. Preliminary human review compares DiscoveryRSI-27B with other research agents.
Unconfirmed: Human comparisons and the 6.3% correspondence rate are preliminary author-reported assessments.
- DiscoveryRSI-27B had a 6.3% rate of correspondence with post-cutoff published research.
- The authors report 2.5% for GPT-6 and 2.3% for Claude Opus 5 on that rate.
- Human raters scored DiscoveryRSI-27B above GPT-6 Luna on importance and evidence-bounded claims.
- Learning uses credit optimization, harness replay, and rediscovery of accepted improvements.
From the original article
1 min read at huggingface.co
Scientific 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, Chenwei Wu, Zhaochen Yu · huggingface.co
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.