DiscoveryRSI is a method study of how research agents can keep improving from their own research experience. The paper first gives an operational definition of a valuable and testable research hypothesis: an evidence-grounded claim about a consequential uncertainty, coupled with a feasible investigation, explicit decision criteria and alternative explanations.
Two things make this hard. Turning a hypothesis into an investigation that can actually run means pinning down its data requirements and verification criteria without changing the question it asks. Learning from research experience means knowing which contributions to reinforce, and a hypothesis's final assessment does not say which step mattered: one agent may have proposed the explanation, another added a useful control, a third revised the plan in response to new evidence. Keep only the final proposal and those contributions disappear.
DiscoveryRSI develops research trees of hypotheses through multi-agent evolutionary search, keeps the evidence and decisions behind every proposal and revision, and compiles selected computational investigations into executable Harbor tasks. These records support three coupled mechanisms. Hypothesis credit optimization attributes assessments of research quality to the documented contributions behind them and uses that to train DiscoveryRSI-27B. Harness replay reads the records for productive strategies and recurring failures and refines the agents' workflow. Rediscovery applies accepted updates to fresh questions, generating the experience the next round learns from. The paper also introduces DFM-Bench, which assesses research quality and progress at the level of hypotheses, revisions, development paths and whole research runs.
The paper evaluates on four life-science domains, clinical and population health, therapeutic target prioritization, disease mechanisms and population multi-omics, and adds a battery setting to test whether the method carries beyond life science. What the agents learn from research experience is mainly how to make an investigation testable: after two rounds of self-improvement, mean hypothesis quality rises from 4.86 to 5.39, with the gains concentrated in testability: falsifiability 4.00 to 6.20, alternative explanations 4.44 to 5.98 and data resolvability 3.73 to 4.99. The share of episodes that develop or repair an existing hypothesis also rises from 77.5% to 91.7%, while the share that restart with a new question falls from 21.7% to 5.0%.
In an expert-supervised human review, 27 reviewers with biomedical or clinical research backgrounds assessed 292 proposals blind. The mean score change per revision is +0.07 for DiscoveryRSI-27B, +0.03 for GPT-5.6 Sol and −0.01 for GPT-6 Luna. Against research published after the model cutoff, 6.3% of DiscoveryRSI-27B's hypotheses have a concrete connection a researcher could build on, such as a usable experimental design, comparator, measurement or testable evidence gap, the highest among the models compared.
Each result has its limits. After two rounds the automated importance and novelty scores fall, so the gains are in how investigations are designed. The human assessments are preliminary, and the revision comparisons are made within research trees. The comparison with published research is descriptive: group sizes and cutoffs differ, and whether researchers adopt these hypotheses needs prospective tracking.
The technical report, the code on GitHub, and the models and dataset on Hugging Face are public, and a DiscoveryRSI live stream is coming soon.