greeky/OPDEvolver

OPD-Evolver

This repository provides PEFT/LoRA adapters for OPD-Evolver, a self-evolving agent framework introduced in the paper OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation.

OPD-Evolver is designed to help agents not only store experience, but also learn how to select useful experience, use it for task solving, write reusable knowledge, and maintain an evolving memory repository.

Released Adapters

This repository contains three adapter modules:

• selector: selects useful experiences from memory.
• executor: uses retrieved experience to solve the current task.
• writer: writes reusable knowledge back into memory.

Note that the experience management module is not included in this version. We will update subsequently.

Paper

For details about the method, training procedure, and evaluation, please refer to:

• Paper: https://arxiv.org/abs/2606.17628
• Project code: https://github.com/bingreeky/opd-evolver

Citation

@misc{zhang2026opdevolver,
  title={OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation},
  author={Guibin Zhang and Xun Xu and Yanwei Yue and Zikun Su and Wangchunshu Zhou and Xiaobin Hu and Shuicheng Yan},
  year={2026},
  eprint={2606.17628},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2606.17628}
}