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

We introduce and release ScienceBuddy, an interactive scientific research workspace that brings continually improving scientific agents into researchers' everyday workflows. ScienceBuddy supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. At its core i

This paper introduces a novel self-improvement paradigm for scientific agents, directly impacting how ML engineers can design and deploy intelligent systems for complex research tasks. Understanding recursive-in-recursive self-improvement is crucial for building robust and continually learning AI assistants, moving beyond static models to dynamic, adaptive agents.