DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
Research that learns from its own research: every step kept in the hypothesis tree becomes experience for the next round.
Move the cursor over the diagram
Overview
Hypothesis discovery that learns recursively from research experience
Multi-agent evolutionary search grows research trees of hypotheses and compiles selected ones into executable tasks; the kept research record trains the model, refines the workflow and is applied to fresh questions.
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 and keeps the evidence and decisions behind every proposal and revision. Selected computational investigations are compiled into executable Harbor tasks with concrete inputs, environments, required outputs and verification checks. The retained research records support three coupled mechanisms. Hypothesis credit optimization attributes assessments of research quality to the documented contributions behind them and turns that into supervision for the associated model actions; the paper uses it to train DiscoveryRSI-27B. Harness replay reads the same records for productive strategies and recurring failures and changes how agents coordinate, access evidence and allocate effort. Rediscovery applies accepted model and harness updates to fresh questions, generating the experience the next round learns from. The paper also introduces DFM-Bench, which evaluates hypotheses, revisions, development paths and research runs, with the execution readiness of compiled tasks recorded separately.
How it is evaluated
- Automated grading (DFM-Bench): hypothesis quality across rounds of self-improvement
- Research behaviour: how the share of each research action changes
- Expert-supervised human review: 27 reviewers assess 292 proposals blind
- Inspiration rate: a comparison with research published after the model cutoff
- A battery setting: whether the method carries beyond life science
The technical report, the code on GitHub, and the models and dataset on Hugging Face are public; a live stream is coming soon.
Abstract
Scientific discovery requires developing worthwhile hypotheses into practical investigations, yet learning from this evolving research process remains challenging. We introduce an operational definition of valuable and testable research hypotheses: evidence-grounded claims about consequential uncertainties, coupled with feasible investigations, explicit decision criteria, and alternative explanations. Our DiscoveryRSI framework develops hypothesis trees through multi-agent evolutionary search, compiles selected investigations into executable Harbor tasks, and recursively learns through three mechanisms: hypothesis credit optimization attributes scientific assessments to documented research contributions; harness replay identifies productive strategies and recurring failures to refine research workflows; and rediscovery applies accepted improvements to new investigations, generating further experience for subsequent learning. We also introduce DFM-Bench, which provides comprehensive assessment of hypothesis quality and cumulative research progress at the levels of hypotheses, revisions, development paths and research runs. Across four life-science domains, automated evaluation shows that hypothesis quality rises from 4.86 for the base model to 5.39 after two rounds of self-improvement, with the large gains in falsifiability and data resolvability. We further validate hypothesis quality through expert review, comparing our DiscoveryRSI-27B with other research agents. Preliminary human assessments show that DiscoveryRSI-27B scores higher than GPT-6 Luna on hypothesis importance and whether claims are plausible and evidence-bounded. DiscoveryRSI-27B achieved a 6.3% rate of hypothesis correspondence with research published after the model cutoff, largely outperform GPT-6 with 2.5% and Claude Opus 5 with 2.3%, while delivering greater human-rated improvement along research paths.
Authors
* Equal contribution
Venue
Cite
BibTeX
@techreport{zhong2026discoveryrsi,
title = {{DiscoveryRSI}: Recursively Learning to Discover Valuable Hypotheses},
author = {Zhong, Jianyuan and Wu, Chenwei and Yu, Zhaochen and Jiang, Xixian and Fan, Jiajun and Li, Xuewei and Xiao, Hao and Dai, Xilin and Wang, Junyang and Zou, Jiaru and Cui, Zhiqing and Zheng, Boyuan and Xiong, Jing and Zhang, Zhen and Gao, Qiang and Morley, Michael and Nakayama, Luis and Xu, Qiang and Shen, Liyue and Gui, Tao and Celi, Leo Anthony and Zou, James and Yin, Zhenfei and Wu, Yingcheng and Yang, Ling},
institution = {PhAI Labs},
type = {Technical Report},
number = {PHAI-TR-2026-06},
month = {October},
year = {2026},
url = {https://phai-labs.com/papers/discoveryrsi/},
note = {Version v1. Equal contribution: Jianyuan Zhong, Chenwei Wu, Zhaochen Yu}
}