EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

Autonomous AI agents are expected to iteratively improve executable policies through feedback, but existing evaluations often reduce this complex process to a single final score, obscuring how the improvement actually happens. EvoPolicyGym introduces a controlled setting where an agent repeatedly edits a policy within compact interactive reinforcement learning environments under a fixed budget. Beyond leaderboard rankings—where GPT-5.5 currently performs best—the benchmark provides detailed trajectory-level diagnostics revealing how agents allocate their effort and refine strategies. This tool is useful for researchers studying and improving how autonomous agents learn and adapt policies over time in RL settings. Authors: Zhilin Wang, Han Song, Runzhe Zhan, Jusen Du, Jiacheng Chen, Tianle Li, Qingyu Yin, Yulun Wu, Zhennan Shen, Tong Zhu, Yanshu Li, Guanjie Chen, Derek F. Wong, Yafu Li, Yu Cheng, Yang Yang Paper: https://arxiv.org/abs/2607.02440v1