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

Author: Geng, Hejia; Huang, Zesen; Li, Haoyang; Li, Wenbin; Wu, Koutian; Zhou, Zihan; Pang, Yuanbo; Liu, Weihao; Xu, Zigong; Li, Zhiping; Zhang, Zongzheng; Dong, Chuanfei; Sun, Jiankai; Zheng, Tianzhe; Xie, Fengyu; Ma, Yue; Shi, Yueheng; Xie, Tong; Di, Zonglin; Liu, Xianrong; Gao, Qucheng; Liu, Yimin; Pan, Jiaming; Huang, Sheng; Ma, Xiao-Han; Yuan, Lanqing; Zhu, Zhenlin; Liu, Ziang; Xu, Ziyang; Wang, Junkai; Liang, Kangkai; Xian, Jiayi; Zhao, Zehong; Xu, Liuwei; Xie, Jingxu; Zhang, Peijin; Gao, Qiang; Xing, Chengyi; Zhao, Zhe; Wang, Xi; Xing, Yaopeng; Meng, Xing; Yin, Zhenfei; Wu, Yingcheng; Yang, Ling.

Published in arXiv (preprint), 2026

Summary: Infrastructure that converts scientific code repositories into executable learning environments with expert-defined cases and scientific verification. Verified interaction trajectories support supervised fine-tuning, reinforcement learning, and evaluation, with reported improvements in held-out scientific-code repair and selected general-purpose tasks.