ScienceIDE trains AI agents on scientific code repositories
Scientific code repositories encode decades of human knowledge in executable models, methods and tools, but fragmented toolchains, implicit domain conventions and specialized correctness criteria make that knowledge hard to convert into reliable learning experience. The authors call this the scientific experience bottleneck.
To address it, they 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 then serve as a shared foundation for supervised fine-tuning, reinforcement learning and evaluation.
Using verified interaction trajectories produced inside these environments, the team trained a family of models: 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, which the authors present as evidence of positive transfer from scientific experience to broader capabilities.
The authors frame ScienceIDE as a foundation for an integrated workspace combining agent learning with scientific practice, treating humanity's scientific software as a shared substrate for developing scientific intelligence. Code is published on GitHub at aitofound/ScienceIDE.
Key facts
- ScienceIDE turns scientific code repositories into programmable, agent-learnable environments, guided by expert-defined scientific cases and acceptance criteria.
- The environments support task generation, execution and scientific verification, and serve as a shared foundation for supervised fine-tuning, reinforcement learning and evaluation.
- The team trained three models on verified interaction trajectories from these environments: PhAI-IDE-72B, PhAI-IDE-9B and PhAI-IDE-4B.
- The PhAI-IDE models show gains in held-out scientific-code repair and on selected general-purpose benchmarks in code, reasoning and knowledge, which the authors read as evidence of positive transfer from scientific to general capabilities.
- Code is published on GitHub at aitofound/ScienceIDE.
Why it matters
Scientific code repositories hold decades of accumulated expertise, but that expertise has been locked behind fragmented toolchains and specialized correctness criteria that generic training pipelines cannot parse. ScienceIDE proposes a systematic way to convert that locked-up code into structured learning material for AI agents, rather than treating scientific software as just another pile of text to train on.
Who it affects
The work targets researchers building AI agents for scientific and technical tasks, and by extension anyone maintaining scientific code repositories that could become training material for such agents. The GitHub release lowers the barrier to applying ScienceIDE's approach to other repositories beyond the ones used in this work.
How to use it
The paper points to a GitHub repository, aitofound/ScienceIDE, as the entry point. The source names three trained models by parameter count, 72B, 9B and 4B, but does not state a license, whether model weights are released, or setup instructions.
How solid is it
The claims rest on the authors' own description of gains in held-out scientific-code repair and in unnamed general-purpose benchmarks covering code, reasoning and knowledge. The source text does not name specific benchmarks or report numeric scores, so the size of the improvement cannot be independently judged from this text alone.
Risks and caveats
The source gives no author names, institutional affiliations, release date, or details on the scale and composition of the training data behind the 'verified interaction trajectories.' The GitHub link is provided without specifying what it contains, whether source code only, the environments themselves, or trained model weights.