ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
Shuhan Xue, Jianyuan Zhong, Ziyuan Nan, Wenbin Li, Zhaochen Yu, Jinchao Ding, Qiang Gao, Pengyu Zhan and 5 others
Published September 15, 2026 · on the feed September 16, 2026 · ▲ 30 · arXiv 2609.17523
The abstract
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
The Brief
2:01A short podcast summary of the paper. Written and voiced by AI.
Transcript
ScienceBuddy. Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents. By Shuhan Xue and twelve others. Published on the fifteenth of September. Its code repository is worked out: about eighty source files, with a readme and a licence.
ScienceBuddy is an interactive scientific research workspace, introduced and released as a research product. Its purpose is to bring continually improving scientific agents into the everyday workflows of researchers, rather than keeping them in a separate lab setting.
It supports researchers as they carry out scientific tasks. Along the way, it transforms what they do into material for learning. Their requests, their feedback, and the evidence from each execution become new tasks and new evaluation rubrics for continual learning. Ordinary use of the workspace is, in effect, also the training signal. A researcher asking for help is at the same time supplying the next thing the agent will be tested on.
At its core is a paradigm the authors call recursive-in-recursive self-improvement, which couples harness evolution with reinforcement learning on the model. There are two loops, one inside the other. The inner recursion improves the harness while the model is held fixed. The outer recursion trains the model under the improved harness. Each feeds the other. Harness evolution shapes the experience the model trains on, and model learning opens new opportunities for the harness to adapt.
The paper presents case studies of researcher interaction, of harness refinement, and of model learning, with benchmark cases spanning four families of scientific task.
It is aimed at the scientific community, and at researchers who want improving agents inside their daily work rather than alongside it. Releasing it as a product is how the authors make the paradigm available to researchers to use directly.
The authors frame it as a step toward what they call discovery intelligence: scientific AI that advances through sustained collaboration with researchers, and evolves alongside the research it supports. The website is science-buddy dot io.
Paper Quest · 7 chapters · about 38 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every correction a researcher makes is a lesson. Most of them are lost when the chat ends.
You are a researcher working with an AI agent. It retrieves evidence and runs analyses; you clarify, question its conclusions and ask for extra checks.
The road ahead
- 1 · The problem ~7 min When researchers correct a scientific agent in a chat, the fix helps that one answer but leaves the agent no better at the next task.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Turn researcher collaboration into scored tasks, then improve the agent's instructions in an inner loop and its model in an outer loop, each feeding the other.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Help researchers, turn their exchanges into tasks with rubrics, refine the harness with the model fixed, then train the model under that harness.
- 6 · What they found ~3 min
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.