Research agents learn from their own research experience: two rounds of self-improvement lift hypothesis testability
Synopsis
Whether research agents can keep improving from their own research experience depends on turning a hypothesis into an investigation that can actually run and on knowing which contribution to reinforce. DiscoveryRSI builds research trees of hypotheses through multi-agent evolutionary search, keeps the evidence and decisions behind every proposal and revision, and uses those records to train the model and refine the workflow. After two rounds of self-improvement, mean hypothesis quality rises from 4.86 to 5.39, with gains concentrated in testability: falsifiability from 4.00 to 6.20 and data resolvability from 3.73 to 4.99.
Interpretation
Self-improvement for research agents no longer rests only on the final proposal; every contribution along the research path is recorded and used for learning. Keeping only the final proposal makes the contributions disappear: one agent proposed the explanation, another added a useful control, a third revised the plan in response to new evidence, and there is no way to tell which step mattered. After two rounds of self-improvement, mean hypothesis quality rises from 4.86 to 5.39, with 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 system compiles selected computational investigations into executable Harbor tasks, so a hypothesis can actually be run. Turning a hypothesis into an investigation that can run means pinning down its data requirements and verification criteria without changing the question it asks. Evaluation covers 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. The share of episodes that develop or repair an existing hypothesis rises from 77.5% to 91.7%, while the share that restart with a new question falls from 21.7% to 5.0%. In a blind review of 292 proposals by 27 reviewers with biomedical or clinical research backgrounds, 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.
Perspective
Evaluation covers four life-science domains, clinical and population health, therapeutic target prioritization, disease mechanisms and population multi-omics, plus a battery setting that tests whether the method carries beyond life science; the technical report, the code on GitHub, and the models and dataset on Hugging Face are public, and the DiscoveryRSI live stream has started.
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.