DiscoveryRSI: self-improving framework for hypothesis-driven research agents

Ling Yang introduces DiscoveryRSI, a framework that runs hypothesis-driven research as a recursive self-improvement loop for research agents.

According to the authors, hypothesis quality on the DFM-Bench benchmark across four life-science domains rose from 4.86 to 5.39 over two rounds. They report that 41% of DiscoveryRSI-27B revisions improved quality, compared with 31% for GPT-5.6 Sol. The authors say importance and novelty dip slightly, and that the model learns to design better tests rather than bolder questions. The code is published at github.com/Gen-Verse/DiscoveryRSI.