HiPO (Hybrid Policy Optimization) was proposed in September 2025 by a research team from Kuaishou and Nanjing University. The relevant research findings were published in the paper "HiPO: Hybrid Policy Optimization for Dynamic Reasoning in LLMs".

HiPO is a framework for adaptive inference control that enables LLMs to selectively decide when to perform detailed inference (Think-on) and when to provide direct responses (Think-off). Specifically, HiPO combines a hybrid data pipeline that provides paired Think-on and Think-off responses with a hybrid reinforcement learning reward system that avoids over-reliance on detailed inference while balancing accuracy and efficiency. Experiments on math and programming benchmarks demonstrate that HiPO can significantly reduce token length while maintaining or improving accuracy.