ScienceBuddy: Recursive-in-Recursive Self-Improvement for Scientific Agents

Shuhan Xue · hf · 2026-09-16

ScienceBuddy is an interactive research workspace that embeds continually improving scientific agents into researchers' workflows, converting requests, feedback, and execution evidence into tasks and eval rubrics. Its core is a recursive-in-recursive paradigm: the inner recursion evolves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Case studies span four scientific task families; released at science-buddy.io.

Original post →

More from coding & agent

coding & agent channel →