DiscoveryRSI

The DiscoveryRSI paper proposes a method that lets a model learn to "discover" problems and continuously improve this ability through a recursive mechanism.

The research team gave an operational definition of a "valuable testable hypothesis" comprising four elements: an evidence-based claim, pointing to a consequential uncertainty, a feasible accompanying research plan, and clear decision criteria and alternative explanations.

The DiscoveryRSI framework consists of three interlocking mechanisms: hypothesis credit optimization, harness replay and rediscovery.

The hypothesis credit optimization mechanism attributes scientific evaluation results to specific research contribution records, enabling the system to locate specific reasoning steps or evidence.

The harness replay mechanism identifies high-yield strategies and recurring failure patterns by replaying the research process, and refines the research workflow accordingly.

The rediscovery mechanism applies accepted improvements to new research tasks and generates new experience for subsequent learning.

Automated evaluation shows that DiscoveryRSI's hypothesis quality rose from 4.86 for the base model to 5.39 after two rounds of self-improvement, with gains concentrated in the falsifiability and data-parseability dimensions.

The correspondence rate with post-cutoff papers is a proxy metric for measuring a model's "foresight," but a high correspondence rate may also come from pattern capture of research hotspots and common method combinations rather than genuine causal insight.

Recursive learning depends on high-quality scientific evaluation signals; if the evaluation itself is biased, recursion may amplify the bias.

The sample size of the expert review is described only as "preliminary," and the hypothesis quality score rising from 4.86 to 5.39 is still far from a "reliable scientific research partner."