@inproceedings{zheng-etal-2026-hmpo,
title = "{HMPO}: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression",
author = "Zheng, Minghui and
Chen, Hongxu and
Ren, Huimin and
Xin, Hongsheng and
Qu, Xiaoyang and
Wang, Ze and
Yang, Shuling and
Peng, Ziyu and
Zhang, Kaike and
Pan, Zhou and
Zhan, Kun",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.885/",
pages = "20418--20433",
ISBN = "979-8-89176-459-0",
abstract = "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** (**H**ybrid **M**edian-length **P**olicy **O**ptimization), 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."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zheng-etal-2026-hmpo">
<titleInfo>
<title>HMPO: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression</title>
</titleInfo>
<name type="personal">
<namePart type="given">Minghui</namePart>
<namePart type="family">Zheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongxu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huimin</namePart>
<namePart type="family">Ren</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongsheng</namePart>
<namePart type="family">Xin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyang</namePart>
<namePart type="family">Qu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ze</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuling</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ziyu</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaike</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhou</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kun</namePart>
<namePart type="family">Zhan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>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** (**H**ybrid **M**edian-length **P**olicy **O**ptimization), 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.</abstract>
<identifier type="citekey">zheng-etal-2026-hmpo</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.885/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>20418</start>
<end>20433</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T HMPO: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression
%A Zheng, Minghui
%A Chen, Hongxu
%A Ren, Huimin
%A Xin, Hongsheng
%A Qu, Xiaoyang
%A Wang, Ze
%A Yang, Shuling
%A Peng, Ziyu
%A Zhang, Kaike
%A Pan, Zhou
%A Zhan, Kun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F zheng-etal-2026-hmpo
%X 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** (**H**ybrid **M**edian-length **P**olicy **O**ptimization), 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.
%U https://aclanthology.org/2026.emnlp-main.885/
%P 20418-20433
Markdown (Informal)
[HMPO: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression](https://aclanthology.org/2026.emnlp-main.885/) (Zheng et al., EMNLP 2026)
ACL
- Minghui Zheng, Hongxu Chen, Huimin Ren, Hongsheng Xin, Xiaoyang Qu, Ze Wang, Shuling Yang, Ziyu Peng, Kaike Zhang, Zhou Pan, and Kun Zhan. 2026. HMPO: Hybrid Median-length Policy Optimization for Chain-of-Thought Compression. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 20418–20433, Budapest, Hungary. Association for Computational Linguistics.