ScienceBuddy paper nests harness evolution inside RL loop

Thirteen researchers led by Shuhan Xue have released ScienceBuddy, an "interactive scientific research workspace" whose central claim is a nested self-improvement loop: an inner recursion that tunes the agent harness while the model is held fixed, wrapped by an outer recursion that then retrains the model under the improved harness.

The abstract calls this "recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning." The authors describe the workspace as capturing researcher requests, feedback, and execution evidence, and turning them into tasks and evaluation rubrics that feed the continual-learning loop. Case studies span "four scientific task families," though the abstract publishes no accuracy figures, baselines, or per-family results.

The framing is explicitly aspirational in the paper's own words. The authors describe ScienceBuddy as "a step toward discovery intelligence: scientific AI that advances through sustained collaboration with researchers and evolves alongside the research it supports." The workspace has been released as a research product rather than a benchmark-only artifact, landing in a busy week of agent-tooling papers on our beat.