Large Language Models (LLMs) increasingly rely on Chain-of-Thought (CoT) reasoning to improve accuracy on complex tasks. However, always generating lengthy reasoning traces is inefficient, leading to excessive token usage and higher inference costs. This paper introduces the Hybrid Policy Optimization (i.e., HiPO), a framework for adaptive reasoning control that enables LLMs to selectively decide when to engage in detailed reasoning (Think-on) and when to respond directly (Think-off). Specifically, HiPO combines a hybrid data pipelineproviding paired Think-on and Think-off responseswith a hybrid reinforcement learning reward system that balances accuracy and efficiency while avoiding over-reliance on detailed reasoning. Experiments across mathematics and coding benchmarks demonstrate that HiPO can substantially reduce token length while maintaining or improving accuracy. Finally, we hope HiPO a can be a principled approach for efficient adaptive reasoning, advancing the deployment of reasoning-oriented LLMs in real-world, resource-sensitive settings.
@article{arxiv.2509.23967,
title = {HiPO: Hybrid Policy Optimization for Dynamic Reasoning in LLMs},
author = {Ken Deng and Zizheng Zhan and Wen Xiang and Wenqiang Zhu and Weihao Li and Jingxuan Xu and Tianhao Peng and Xinping Lei and Kun Wu and Yifan Yao and Haoyang Huang and Huaixi Tang and Kepeng Lei and Zhiyi Lai and Songwei Yu and Zongxian Feng and Zuchen Gao and Weihao Xie and Chenchen Zhang and Yanan Wu and Yuanxing Zhang and Lecheng Huang and Yuqun Zhang and Jie Liu and Zhaoxiang Zhang and Haotian Zhang and Bin Chen and Jiaheng Liu},
journal= {arXiv preprint arXiv:2509.23967},
year = {2025}
}