ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments
Hejia Geng, Zesen Huang, Haoyang Li, Wenbin Li, Koutian Wu, Zihan Zhou, Yuanbo Pang, Weihao Liu and 37 others
AItonomy Foundation · University of Oxford · University of California, Los Angeles · University of Texas at Austin · Princeton University · University of California, Berkeley
Published September 16, 2026 · on the feed September 17, 2026 · ▲ 100 · arXiv 2609.19134
The abstract
Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain conventions, and specialized correctness criteria make this knowledge difficult to convert into reliable learning experience-a challenge we call the scientific experience bottleneck. We introduce ScienceIDE, infrastructure for turning the world's scientific code into programmable environments for scientific agents. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments that support task generation, execution, and scientific verification. These environments provide a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories, we train PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities. ScienceIDE lays the foundation for an integrated workspace for agent learning and scientific practice, making humanity's scientific software a shared substrate for developing scientific intelligence. Code: https://github.com/aitofound/ScienceIDE
The Brief
2:29A short podcast summary of the paper. Written and voiced by AI.
Transcript
ScienceIDE. Turning World's Scientific Codebase into Agent Learnable Environments. By Hejia Geng and forty-four others, from the Aitonomy Foundation, the University of Oxford, and twenty-three other institutions. Published on the sixteenth of September. Its code repository is worked out: about one thousand five hundred source files, with a readme.
Scientific code repositories hold decades of human knowledge in executable form: models, methods and tools. Turning that knowledge into reliable learning experience for agents is hard. Toolchains are fragmented, domain conventions are implicit, and correctness criteria are specialized. The authors call this the scientific experience bottleneck. ScienceIDE is infrastructure for getting past it, by turning the world's scientific code into programmable environments for scientific agents. Much scientific knowledge lives in code rather than in papers, and code, unlike prose, can be run. The aim is to make that code something an agent can practise on.
Each environment is an executable form of a repository that supports three things: generating tasks, executing them, and verifying the results scientifically. Together the environments form a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. An environment, in this sense, is a place where an agent can be set a scientific task, attempt it by running real scientific code, and have its result checked against scientific standards rather than only against whether the code ran.
The environments are built by agents, guided by scientific cases and acceptance criteria that experts define. The agents transform repositories into environments, and the expert criteria are what the scientific verification checks against. Interaction trajectories that pass verification then become training data. So the experts set the standard, and the agents do the conversion work at scale.
The first users are the authors' own models. Using the verified trajectories, they train a family of three: PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The family improves on held-out repair of scientific code, and across selected general-purpose benchmarks in code, reasoning and knowledge. The authors read this as evidence of positive transfer, from scientific experience to broader capabilities. The code is public.
The ambition is an integrated workspace for agent learning and scientific practice, one that makes humanity's scientific software a shared substrate for developing scientific intelligence. The first result suggests the experience gained there does not stay inside science.
Paper Quest · 7 chapters · about 40 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
The world's science is written in code. Why can't agents learn from it yet?
You want to train agents on real scientific work, and decades of knowledge already sit in scientific code repositories. Unlike prose, code can be run, so an agent could practise on it.
The road ahead
- 1 · The problem ~6 min Scientific code holds decades of runnable knowledge, but a repository gives an agent no tasks, no feedback and no way to judge scientific correctness.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Experts define what counts as scientifically correct once; agents turn repositories into environments that generate, run and verify tasks at scale.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Experts set checks, agents build environments, tasks are generated and verified, verified trajectories are kept, and the PhAI-IDE models are fine-tuned on them.
- 6 · The experiment ~6 min You’ll guess: After training on verified scientific-code trajectories, what happened to the PhAI-IDE models on general code, reasoning and knowledge benchmarks?
- 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.