ScienceBuddy

ScienceBuddy is an interactive research workspace that pairs a scientific AI agent with a working researcher. Requests, feedback, and execution evidence become new tasks and evaluation rubrics that the system learns from. It comes from "ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents" (arXiv:2609.17523), posted 15 September 2026 by Shuhan Xue, Ling Yang, and 11 other authors.

Its mechanism, recursive-in-recursive self-improvement, nests two update loops instead of one. An inner recursion edits the agent's harness, its scaffolding and tools, while holding the model fixed. An outer recursion then trains that model with reinforcement learning under the improved harness. Harness changes shape what the model experiences during training, and what the model learns opens new options for harness edits, so the two loops feed each other. This makes ScienceBuddy a specific two-level case of the broader recursive self-improvement idea already catalogued in this vocabulary as RSI, not a new general framework.

The paper reports case studies of researcher interaction, harness refinement, and model learning, with benchmark tasks drawn from four scientific task families. The authors frame the project as a step toward what they call "discovery intelligence." ScienceBuddy is released as a research product with a project site at science-buddy.io and a code repository on GitHub. The task families and the discovery-intelligence framing come from the paper's 13 authors and are not independently verified.