Researchers introduced DiscoveryRSI in a paper published on October 16 to generate testable scientific hypotheses through multi-agent evolutionary search and recursive learning. The framework defines a valuable hypothesis as an evidence-grounded claim about an important uncertainty, paired with a feasible investigation, explicit decision criteria and alternative explanations. It builds hypothesis trees, turns selected investigations into executable Harbor tasks and learns from the research process.

DiscoveryRSI uses three learning mechanisms: hypothesis credit optimization assigns scientific assessments to documented research contributions; harness replay finds productive strategies and recurring failures to refine workflows; and rediscovery applies accepted improvements to new investigations to create further learning experience. The authors also introduce DFM-Bench, which assesses hypothesis quality and cumulative progress across hypotheses, revisions, development paths and research runs.

Across four life-science domains, automated evaluation scores rose from 4.86 for the base model to 5.39 after two rounds of self-improvement, with the largest gains in falsifiability and data resolvability. Expert review compared DiscoveryRSI-27B with other research agents. Preliminary human assessments rated it higher than GPT-6 Luna for hypothesis importance and whether claims were plausible and bounded by evidence. DiscoveryRSI-27B had a 6.3% correspondence rate with hypotheses in research published after its model cutoff, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5; it also showed greater human-rated improvement along research paths.

Researchers introduced DiscoveryRSI in a paper published on October 16 to generate testable scientific hypotheses through multi-agent evolutionary search and recursive learning. The framework defines a valuable hypothesis as an evidence-grounded claim about an important uncertainty, paired with a feasible investigation, explicit decision criteria and alternative explanations. It builds hypothesis trees, turns selected investigations into executable Harbor tasks and learns from the research process.

DiscoveryRSI uses three learning mechanisms: hypothesis credit optimization assigns scientific assessments to documented research contributions; harness replay finds productive strategies and recurring failures to refine workflows; and rediscovery applies accepted improvements to new investigations to create further learning experience. The authors also introduce DFM-Bench, which assesses hypothesis quality and cumulative progress across hypotheses, revisions, development paths and research runs.

Across four life-science domains, automated evaluation scores rose from 4.86 for the base model to 5.39 after two rounds of self-improvement, with the largest gains in falsifiability and data resolvability. Expert review compared DiscoveryRSI-27B with other research agents. Preliminary human assessments rated it higher than GPT-6 Luna for hypothesis importance and whether claims were plausible and bounded by evidence. DiscoveryRSI-27B had a 6.3% correspondence rate with hypotheses in research published after its model cutoff, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5; it also showed greater human-rated improvement along research paths.