ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments

ScienceIDE transforms scientific code repositories into executable, verifiable agent training environments, yielding the PhAI-IDE model family at 72B, 9B, and 4B scales with gains on scientific-code repair and general benchmarks.

ScienceIDE is infrastructure that turns scientific code repositories into executable agent-learnable environments supporting task generation, execution, and scientific verification, guided by expert-defined scientific cases and acceptance criteria. It addresses the 'scientific experience bottleneck' arising from fragmented toolchains and implicit conventions. Using verified interaction trajectories from the transformed repositories, the authors train three models: PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The family shows gains on held-out scientific-code repair tasks and positive transfer to selected general-purpose code, reasoning, and knowledge benchmarks. The environments support supervised fine-tuning, reinforcement learning, and evaluation, and the code is released on GitHub.

• ScienceIDE converts scientific code repositories into executable environments for task generation, execution, and scientific verification
• Environments are guided by expert-defined scientific cases and acceptance criteria
• The work targets the 'scientific experience bottleneck' caused by fragmented toolchains and implicit conventions (arXiv report)
• Three models are trained from verified agent interaction trajectories: PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B
• Gains reported on held-out scientific-code repair tasks
• Positive transfer shown to selected general code, reasoning, and knowledge benchmarks
• Environments support supervised fine-tuning, reinforcement learning, and evaluation
• Code is released on GitHub (arXiv report)