‹ back to 2026-09-17

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

arXiv cs.CL LLM Agents Hejia Geng, Zesen Huang, Haoyang Li, Wenbin Li, Koutian Wu, Zihan Zhou, Yuanbo Pang, Weihao Liu, Zigong Xu, Zhiping Li, Zongzheng Zhang, Chuanfei Dong, Jiankai Sun, Tianzhe Zheng, Fengyu Xie, Yue Ma, Yueheng Shi, Tong Xie, Zonglin Di, Xianrong Liu, Qucheng Gao, Yimin Liu, Jiaming Pan, Sheng Huang, Xiao-Han Ma, Lanqing Yuan, Zhenlin Zhu, Ziang Liu, Ziyang Xu, Junkai Wang, Kangkai Liang, Jiayi Xian, Zehong Zhao, Liuwei Xu, Jingxu Xie, Peijin Zhang, Qiang Gao, Chengyi Xing, Zhe Zhao, Xi Wang, Yaopeng Xing, Xing Meng, Zhenfei Yin, Yingcheng Wu, Ling Yang 2026-09-16

TL;DR - ScienceIDE converts scientific code repositories into executable, verifiable environments for training and evaluating AI agents. Models trained on verified scientific interaction trajectories improved on held-out code repair and selected broader benchmarks, suggesting positive transfer.

Expert-defined cases and acceptance criteria guide repository transformation, task generation, execution, and scientific verification.
The environments support supervised fine-tuning, reinforcement learning, and evaluation through verified agent trajectories.
The authors trained PhAI-IDE models at 72B, 9B, and 4B parameter scales.
Reported gains span held-out scientific-code repair and selected general-purpose code, reasoning, and knowledge benchmarks.

view merged work →