Paper Overview
Field: NLP Authors: Jianyuan Zhong, Chenwei Wu, Zhaochen Yu, Xixian Jiang, Xuewei Li, Hao Xiao, Xilin Dai, Junyang Wang, Jiaru Zou, Zhiqing Cui, Boyuan Zheng, Jing Xiong, Zhen Zhang, Qiang Gao, Michael Morley, Luis Nakayama, Qiang Xu, Liyue Shen, Tao Gui, Leo Anthony Celi, James Zou, Zhenfei Yin, Yingcheng Wu, Ling Yang Published: 2026-10-16 arXiv: 2610.23839
Summary
Scientific discovery requires developing worthwhile hypotheses into practical investigations, yet learning from this evolving research process remains challenging. The paper introduces 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.
The DiscoveryRSI Framework
DiscoveryRSI 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.
- Rediscovery: applies accepted improvements to new research, generating further experience for subsequent learning.
- In automated evaluation across four life-science domains, hypothesis quality improved from a baseline model score of 4.86 to 5.39 after two rounds of self-improvement, with the largest gains in falsifiability and data solvability.
- Expert review further validated hypothesis quality: DiscoveryRSI-27B outperformed GPT-6 Luna on hypothesis importance and on whether claims are reasonable and evidence-supported.
- On hypothesis correspondence with research published after the models' knowledge cutoff, DiscoveryRSI-27B achieved 6.3%, substantially outperforming GPT-6 (2.5%) and Claude Opus 5 (2.3%).
DFM-Bench
The authors also release DFM-Bench, a benchmark that comprehensively evaluates hypothesis quality and cumulative research progress across four levels: hypotheses, revisions, development paths, and research runs.
Key Results
Original 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 resea... *(excerpted)*
---
*Auto-collected on 2026-10-20*