ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
Source: https://arxiv.org/abs/2609.17523 · platform: arxiv · 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 · date: 2026-09-15 · meta: Website: http://science-buddy.io, Code: https://github.com/Gen-Verse/ScienceBuddy-RSI
TL;DR（J(中)文摘要）
ScienceBuddy \u662f\u4e00\u4e2a\u5f00\u653e\u7684\u4ea4\u4e92\u5f0f\u79d1\u7814 workspace\uff0c\u628a\u300c\u6301\u7eed\u8fdb\u5316\u7684\u79d1\u7814 agent\u300d\u5d4c\u8fdb\u7814\u7a76\u8005\u65e5\u5e38\u5de5\u4f5c\u6d41\uff1a\u628a\u8bf7\u6c42\u3001\u53cd\u9988\u3001\u6267\u884c\u8bc1\u636e\u7f16\u8bd1\u6210\u4efb\u52a1\u4e0e\u8bc4\u4f30 rubric \u505a\u6301\u7eed\u5b66\u4e60\u3002\u6838\u5fc3\u662f recursive-in-recursive self-improvement\u2014\u2014\u5185\u5c42 recursion \u5728\u6a21\u578b\u51bb\u7ed3\u65f6\u8fdb\u5316 harness\uff0c\u5916\u5c42 recursion \u5728\u65b0 harness \u4e0b\u8bad\u7ec3\u6a21\u578b\uff1bharness \u8fdb\u5316\u5851\u5f62\u8bad\u7ec3\u7ecf\u9a8c\uff0c\u6a21\u578b\u5b66\u4e60\u53cd\u8fc7\u6765\u4e3a harness \u8fed\u4ee3\u5f00\u65b0\u673a\u4f1a\u3002\u8bba\u6587\u5c55\u793a\u56db\u7c7b\u79d1\u7814\u4efb\u52a1\u7684 case study\uff0c\u5e76\u53d1\u5e03\u7814\u7a76\u4ea7\u54c1 science-buddy.io \u63a8\u52a8 discovery intelligence\u3002
Summary (English)
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 is recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning: the inner recursion improves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Harness evolution shapes training experience, and model learning creates new opportunities for harness adaptation. We present case studies of researcher interaction, harness refinement, and model learning, with the benchmark cases spanning four scientific task families. By releasing ScienceBuddy as a research product, we make this paradigm available to the scientific community and take a step toward discovery intelligence: scientific AI that advances through sustained collaboration with researchers and evolves alongside the research it supports. Website: http://science-buddy.io
入库依据（OpenCLI arXiv 直出）
opencli arxiv paper 2609.17523 -f json \u5b8c\u6574 metadata \u89e3\u6790\uff08authors 13 \u4eba\u3001cs.AI/cs.CL\u3001published/updated \u540c\u65e5 2026-09-15\uff09\uff0cabstract \u5373 content \u4e3b\u4f53\uff1bcomment \u7ed9\u51fa\u5b98\u7f51\u4e0e\u4ee3\u7801\u4ed3\u5e93\u94fe\u63a5\u3002