ScienceBuddy paper ships recursive-in-recursive agent workspace
TL;DR
- Thirteen authors released ScienceBuddy on arXiv as an interactive scientific research workspace that turns researcher use into training signal.
- Its 'recursive-in-recursive' design has an inner loop that refines the agent harness with the model fixed, and an outer loop that trains the model under the improved harness.
- Case studies span four scientific task families; the abstract names no benchmarks and reports no per-task numbers.
Thirteen researchers released ScienceBuddy on arXiv, an interactive scientific research workspace that pairs an agent harness with a model that trains on the workspace's own use.
The paper's core claim is a paradigm its authors call "recursive-in-recursive self-improvement." An inner loop refines the harness while the model stays fixed; an outer loop then trains the model under the improved harness. "Harness evolution shapes training experience, and model learning creates new opportunities for harness adaptation," the abstract states.
The system ingests researchers' "requests, feedback, and execution evidence" and converts them into tasks and evaluation rubrics for continual learning. Case studies in the paper span four scientific task families. The abstract names no benchmarks and publishes no per-task numbers.
The authors frame the release as "a step toward discovery intelligence: scientific AI that advances through sustained collaboration with researchers." Code and a hosted workspace at science-buddy.io accompany the paper.
Originally reported by paper
Read the original article →Original headline: ScienceBuddy Ships Open-Source Dual-Loop RSI Platform for Scientific Research