Researchers introduce ScienceIDE, infrastructure that converts scientific code repositories into programmable, verifiable environments for training and evaluating scientific agents, addressing what the authors term the ‘scientific experience bottleneck’ created by fragmented toolchains and implicit domain conventions. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments supporting task generation, supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories from this infrastructure, the team trains a PhAI-IDE model family (72B/9B/4B) that shows gains on held-out scientific code repair as well as positive transfer to general-purpose code, reasoning, and knowledge benchmarks.
