DSPO: Stable and Efficient Policy Optimization for Agentic Search and Reasoning

Official repository for the paper:

DSPO: Stable and Efficient Policy Optimization for Agentic Search and Reasoning

Paper

• arXiv: https://arxiv.org/abs/2510.09255
• Title: DSPO: Stable and Efficient Policy Optimization for Agentic Search and Reasoning
• Authors: Chenyang Gu, Yewen Pu, Bruce Yang, Xiaofan Li, Huan Gao
• Version: arXiv v4, revised on March 19, 2026

Overview

Dynamic-filter Sequence-level Policy Optimization (DSPO) is an RL algorithm designed for stable and efficient agentic search and reasoning.

DSPO trains models to interleave multi-turn search and reasoning through reinforcement learning, using sequence-level optimization and dynamic sample filtering to improve training stability and performance.

News

• [2026-03-19] Paper updated to arXiv v4.
• [2025-10-10] Paper first released on arXiv.

Repository Status

Code and training details will be released soon.

Citation

If you find this work useful, please cite:

@article{gu2025dspo,
  title={DSPO: Stable and Efficient Policy Optimization for Agentic Search and Reasoning},
  author={Gu, Chenyang and Pu, Yewen and Yang, Bruce and Li, Xiaofan and Gao, Huan},
  journal={arXiv preprint arXiv:2510.09255},
  year={2025}
}

License

The paper and figures are licensed under CC BY-SA 4.0.

Code, if released, will be licensed separately.