Can research agents get better at science by learning from the research they do?
Introducing DiscoveryRSI 🧬: a framework that turns hypothesis-driven research into a recursive self-improvement loop.
📄 Paper: https://arxiv.org/abs/2610.23839
💻 Code: https://github.com/Gen-Verse/DiscoveryRSI
🔴 Live Arena: https://foundry-arena.rsi-arena.workers.dev/arena
🔹 What is a valuable hypothesis?
We define it as a joint object H = (claim, investigation plan, decision criteria, alternative explanations). A hypothesis is valuable if resolving it would change a research decision. It is testable if the plan can produce evidence that supports, weakens, or bounds the claim.
🔹 How it works
- Lead, Generation, Evolution, Critique and Progress agents grow hypothesis trees through evolve / repair / crossover / restart
- Every version stays linked to its evidence, data, assessments and contribution records
- Selected plans compile into executable Harbor tasks
🔹 Three self-improvement mechanisms
① Hypothesis credit optimization: blinded TrueSkill assessments assign bounded credit to the agent steps that actually contributed, and DiscoveryRSI-27B is trained only on those steps (Qwen3.8-27B init, 256K context; 2,397 of 15,792 trajectories kept; 2.12% of tokens supervised)
② Harness replay: diagnose recurring failures such as repeated exploration, failed tool use, lost artifacts and premature closure, then refine the workflow
③ Rediscovery: run the accepted model and harness updates on fresh questions
🔹 Results (DFM-Bench, 4 life-science domains)
- Hypothesis quality 4.86 → 5.18 → 5.39 over two rounds
- Falsifiability 4.00 → 6.20, alternative explanations 4.44 → 5.98, data resolvability 3.73 → 4.99
- Evolve/repair share 77.5% → 91.7%, restarts 21.7% → 5.0%: the agent learns to keep developing an idea instead of starting over
- Caveat: importance and novelty dip slightly. The model learns to design better tests, not yet to choose bolder questions.
🔹 Expert review
27 physicians and bioscience researchers scored 292 anonymized proposals:
- 41% of DiscoveryRSI-27B revisions improve quality vs 31% (GPT-5.6 Sol) and 25% (GPT-6 Luna)
- Only 12% make things worse vs 32% and 31%
🔹 Does it anticipate real science?
Against 16,440 publications filtered by each model's cutoff, 6.3% of DiscoveryRSI-27B's hypotheses connect to later published research (16.9% in therapeutic targets). Comparison: DeepSeek 5.4%, GLM 2.8%, GPT 2.5%, Claude 2.3%.
Watch six models do research in real time in DiscoveryRSI Arena, and follow every proposal, critique and compile step 👇
https://foundry-arena.rsi-arena.workers.dev/arena
Co-led by Jianyuan Zhong, Chenwei Wu & Zhaochen Yu, with wonderful collaborators across PhAI Labs, CUHK, Stanford, MIT, Harvard Medical School, Oxford, Princeton, PKU and more.
Grateful to @james_y_zou @Jiaru_Zou @Charles_Y_Wu @zhenfei_yin_yzf 🙏