DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
TLDR
Introduces DiscoveryRSI, a multi-agent framework that recursively learns to propose valuable hypotheses, with a benchmark DFM-Bench for evaluating hypothesis quality across life-science domains.
Reasoning
The paper presents a novel framework combining evolutionary search, recursive learning mechanisms, and a new benchmark, with empirical validation across multiple life-science domains. Weaknesses include limited visible details on the benchmark construction and potential over-reliance on automated evaluation, though expert review is mentioned.
Read-first score
Read-first score 62.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 63.
Field roles
Rank sensitivity
Stability: volatile; rank range: 48.
Keyword Scores
Deep Analysis
Innovations
- Operational definition of valuable and testable research hypotheses as evidence-grounded claims about consequential uncertainties, paired with feasible investigations, explicit decision criteria, and alternative explanations.
- DiscoveryRSI framework that develops hypothesis trees through multi-agent evolutionary search and compiles selected investigations into executable Harbor tasks.
- Recursive learning mechanisms: hypothesis credit optimization, harness replay, and rediscovery.
- DFM-Bench benchmark for assessing hypothesis quality and cumulative research progress at the levels of hypotheses, revisions, development paths, and research runs.
Methodology
DiscoveryRSI uses multi-agent evolutionary search to build hypothesis trees, compiles selected investigations into executable Harbor tasks, and recursively learns through hypothesis credit optimization, harness replay, and rediscovery. Evaluation uses DFM-Bench across four life-science domains and compares DiscoveryRSI-27B with baselines including GPT-6 Luna, GPT-6, and Claude Opus 5 using automated metrics and expert review.
Key Results
Automated evaluation shows hypothesis quality improves from 4.86 for the base model to 5.39 after two rounds of self-improvement, with large gains in falsifiability and data resolvability. Preliminary human assessment finds DiscoveryRSI-27B scores higher than GPT-6 Luna on hypothesis importance and plausibility/evidence-boundedness, achieves 6.3% correspondence with post-cutoff published research versus 2.5% for GPT-6 and 2.3% for Claude Opus 5, and shows greater human-rated improvement along research paths.
Limitations
- Human assessments are preliminary.
- Main quality gains are demonstrated through automated evaluation, with expert validation still preliminary.
- Evaluation is limited to four life-science domains.
- Reported self-improvement results cover two rounds of learning.