Saved in:
Bibliographic Details
Main Authors: Zheng, Minghui, Chen, Hongxu, Ren, Huimin, Xin, Hongsheng, Qu, Xiaoyang, Wang, Ze, Yang, Shuling, Peng, Ziyu, Zhang, Kaike, Zhou, Pan, Zhan, Kun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2606.01934
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911739774238720
author Zheng, Minghui
Chen, Hongxu
Ren, Huimin
Xin, Hongsheng
Qu, Xiaoyang
Wang, Ze
Yang, Shuling
Peng, Ziyu
Zhang, Kaike
Zhou, Pan
Zhan, Kun
author_facet Zheng, Minghui
Chen, Hongxu
Ren, Huimin
Xin, Hongsheng
Qu, Xiaoyang
Wang, Ze
Yang, Shuling
Peng, Ziyu
Zhang, Kaike
Zhou, Pan
Zhan, Kun
contents Large language models achieve remarkable performance via extended chain-of-thought (CoT) reasoning, yet this lengthy process incurs substantial inference overhead. Existing CoT compression methods struggle with inflexible manual length budgets, computationally expensive multi-stage training pipelines, and fragile scalability restricted to small models. We propose HMPO (Hybrid Median-length Policy Optimization), a cost-effective, single-stage reinforcement learning framework. HMPO efficiently compresses CoT via three synergistic components: an adaptive median-based budget derived from successful rollouts to eliminate manual tuning, a cosine-decay token reward for smooth length penalization, and a multiplicative reward formulation that substantially mitigates trivial reward hacking by strictly prioritizing answer correctness. Trained exclusively on mathematical data, HMPO generalizes seamlessly across math, code, science, and instruction-following tasks. Extensive experiments scaling from 9B to 122B parameters across dense and Mixture-of-Experts (MoE) architectures demonstrate that HMPO achieves 19%--46% token compression with negligible accuracy degradation, all while drastically reducing training costs compared to existing multi-stage baselines.
format Preprint
id arxiv_https___arxiv_org_abs_2606_01934
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle HMPO: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression
Zheng, Minghui
Chen, Hongxu
Ren, Huimin
Xin, Hongsheng
Qu, Xiaoyang
Wang, Ze
Yang, Shuling
Peng, Ziyu
Zhang, Kaike
Zhou, Pan
Zhan, Kun
Machine Learning
Computation and Language
Large language models achieve remarkable performance via extended chain-of-thought (CoT) reasoning, yet this lengthy process incurs substantial inference overhead. Existing CoT compression methods struggle with inflexible manual length budgets, computationally expensive multi-stage training pipelines, and fragile scalability restricted to small models. We propose HMPO (Hybrid Median-length Policy Optimization), a cost-effective, single-stage reinforcement learning framework. HMPO efficiently compresses CoT via three synergistic components: an adaptive median-based budget derived from successful rollouts to eliminate manual tuning, a cosine-decay token reward for smooth length penalization, and a multiplicative reward formulation that substantially mitigates trivial reward hacking by strictly prioritizing answer correctness. Trained exclusively on mathematical data, HMPO generalizes seamlessly across math, code, science, and instruction-following tasks. Extensive experiments scaling from 9B to 122B parameters across dense and Mixture-of-Experts (MoE) architectures demonstrate that HMPO achieves 19%--46% token compression with negligible accuracy degradation, all while drastically reducing training costs compared to existing multi-stage baselines.
title HMPO: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2606.01934