CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution

Shidong Yang*, Ziyu Ma*, Tongwen Huang*, Yiming Hu, Yong Wang†, Xiangxiang Chu
AMAP, Alibaba Group
* Equal contribution. † Project lead and corresponding author.

News

• [2026.04]: 🚀 SkillClaw released — a real-environment collective skill evolution system where reusable agent skills evolve from every interaction across sessions, agents, devices, and users, allowing distributed experience to compound into collective intelligence.
• [2026.04]: 🎉 Our paper is accepted to ACL 2026

Table of contents

• Overview
• Acknowledgement
• Citation
• License

Overview

CoEvolve studies reinforcement learning for LLM agents under a changing training distribution. Instead of relying on a fixed pool of expert demonstrations or static synthetic trajectories, CoEvolve closes the loop between the agent and its data: the current policy interacts with the environment, failure signals are extracted from rollouts, and those signals guide the synthesis of new tasks that are validated and folded back into training.

Acknowledgement

This codebase is developed on top of AgentEvolver and veRL. Some components are inspired by CuES. We sincerely thank the authors and contributors of these open-source projects.

Citation

If you find this repository useful in your research, please cite the arXiv version below:

@article{yang2026coevolve,
  title={CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution},
  author={Shidong Yang and Ziyu Ma and Tongwen Huang and Yiming Hu and Yong Wang and Xiangxiang Chu},
  journal={https://arxiv.org/abs/2604.15840},
  year={2026}
}

License

Apache-2.0. See LICENSE for details.