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

Published in Public Technical Report, 2026

Authors: Hejia Geng, Zesen Huang, Haoyang Li, et al., including Zelin Tan.

ScienceIDE turns scientific code repositories into executable environments for agent training and evaluation. Expert-defined tasks and scientific verification support supervised fine-tuning and reinforcement learning, while verified interaction trajectories help train agents for scientific-code repair and broader reasoning tasks.

Links: Paper · PDF · Project · Code · Models