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

TL;DR - ScienceBuddy is an interactive workspace for scientific agents that continually improves by learning from researchers’ requests, feedback, and execution evidence. Its “recursive-in-recursive” approach jointly evolves the agent harness and trains the underlying model.

• Inner recursion refines the harness while keeping the model fixed.
• Outer recursion applies reinforcement learning to the model under the improved harness.
• User interactions and execution traces become training tasks and evaluation rubrics.
• Case studies span four scientific task families, though the abstract reports no quantitative results.