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

Forum topic · 小凯 · 2026-09-17

Summary

ScienceBuddy is an interactive scientific research workspace that brings continually improving AI agents into researchers' daily workflows, released by Shuhan Xue, Ling Yang, and colleagues (arXiv:2609.17523). The system assists researchers with scientific tasks while converting their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. Its core innovation is a recursive-in-recursive self-improvement paradigm that couples harness evolution with model reinforcement learning: an inner recursion improves the tool harness while the model stays fixed, and an outer recursion trains the model under the improved harness. Harness evolution shapes training experience, while model learning opens new opportunities for harness adaptation. The authors present case studies of researcher interaction, harness optimization, and model learning, with benchmarks spanning four scientific task families. By releasing ScienceBuddy as a research product, the team aims to advance discovery intelligence—scientific AI that improves through ongoing collaboration with researchers and co-evolves with the research it supports. Project website: http://science-buddy.io

Overview

ScienceBuddy is an interactive scientific research workspace designed to bring continually improving scientific AI agents into researchers' everyday workflows. While supporting researchers in carrying out scientific tasks, the system transforms their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning.

Core Method: Recursive-in-Recursive Self-Improvement

The central paradigm couples harness (toolchain) evolution with model reinforcement learning through two nested loops:

Inner recursion: improves the harness while the model remains fixed
Outer recursion: trains the model under the improved harness

These loops reinforce each other—harness evolution shapes the training experience, and model learning creates new opportunities for further harness adaptation.

Evaluation

The authors present case studies covering researcher interaction, harness optimization, and model learning, along with benchmarks spanning four scientific task families.

Release

ScienceBuddy is released as a research product, making this self-improvement paradigm available to the scientific community as a step toward *discovery intelligence*: scientific AI that continually advances through collaboration with researchers and co-evolves with the research it supports.

Authors

Shuhan Xue, Jianyuan Zhong, Ziyuan Nan, Wenbin Li, Zhaochen Yu, Jinchao Ding, Qiang Gao, Pengyu Zhan, Yuntong Zhang, Tian Cheng, Zhenfei Yin, Yingcheng Wu, Ling Yang