DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

DiscoveryRSI. Recursively Learning to Discover Valuable Hypotheses. By Jianyuan Zhong and twenty-three others, from PhAI Labs, the Chinese University of Hong Kong, and fourteen other institutions. Published on the sixteenth of October. Its code repository is public and real: thirteen source files.

Scientific discovery means taking hypotheses that deserve attention and developing them into practical investigations. Learning from that process is hard, because the process keeps evolving as the research goes on: hypotheses get revised and investigations change direction along the way. Part of the difficulty is simply definitional. Before a system can learn to produce valuable hypotheses, someone has to say what a valuable hypothesis is.

The paper begins there, with an operational definition. A valuable, testable research hypothesis is an evidence-grounded claim about a consequential uncertainty. It comes coupled with a feasible investigation, explicit decision criteria, and alternative explanations. The DiscoveryRSI framework is built to produce hypotheses of that kind, and to get better at producing them, recursively, from its own experience.

It develops hypothesis trees through multi-agent evolutionary search, then compiles the selected investigations into executable Harbor tasks. Learning happens through three mechanisms. Hypothesis credit optimization attributes scientific assessments back to the documented research contributions that earned them. Harness replay identifies productive strategies and recurring failures, and uses both to refine research workflows. Rediscovery applies the accepted improvements to new investigations, which generates fresh experience for the next round of learning, closing the loop.

To measure progress, the authors also introduce DFM-Bench. It provides a comprehensive assessment of hypothesis quality and of cumulative research progress, at four levels: single hypotheses, revisions, development paths, and whole research runs.

Across four life-science domains, automated evaluation shows hypothesis quality rising from four point eight six for the base model to five point three nine after two rounds of self-improvement. The largest gains are in falsifiability and data resolvability. Expert review then compares DiscoveryRSI-27B with other research agents. Preliminary human assessments rate it above GPT-6 Luna on hypothesis importance, and on whether claims are plausible and evidence-bounded.

The most concrete measure looks forward in time. DiscoveryRSI-27B reached a six point three percent rate of correspondence with research published after its model cutoff. GPT-6 reached two point five percent, and Claude Opus 5 reached two point three percent. Human raters also judged it to deliver greater improvement along research paths.

Strip away the machinery and the loop is familiar from any lab. Propose, investigate, assess, credit what worked, and carry that into the next project. DiscoveryRSI writes each of those steps down as something a model can learn from, starting with a definition of what a good hypothesis is. The human assessments are labelled preliminary, and that label belongs with the numbers.