@inproceedings{gao-etal-2025-towards,
title = "Towards A Better Initial Policy Model For Scalable Long-{C}o{T} Reinforcement Learning",
author = "Gao, Bofei and
Wang, Yejie and
Miao, Yibo and
Wu, Ruoyu and
Song, Feifan and
Yu, Longhui and
Liu, Tianyu and
Chang, Baobao",
editor = "Che, Wanxiang and
Nabende, Joyce and
Shutova, Ekaterina and
Pilehvar, Mohammad Taher",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-acl.397/",
doi = "10.18653/v1/2025.findings-acl.397",
pages = "7652--7665",
ISBN = "979-8-89176-256-5",
abstract = "Long-CoT reasoning combined with reinforcement learning for large language models demonstrates remarkable performance and scalability. However, we observe that the initial policy model could significantly influence the final performance as well as the token efficiency. Additionally, there is a lack of systematic guidelines for obtaining a better initial policy model. To bridge this gap, we initiate a comprehensive investigation by activating the initial model using a variety of datasets with different data volumes and reasoning patterns. Then, we conduct a thorough analysis and comparison of the RL process for different initial models from the perspectives of upper bounds, diversity, and token efficiency, providing a deeper understanding and insight into the long-CoT RL. Based on our empirical results, we propose a systematic guideline and a novel Re-RFT method for constructing a better RL start point. Our experiment results based on the 14B model surpass the DeepSeek-R1-Distill-Qwen-14B by an average of 4.6{\%}, demonstrating our approach{'}s effectiveness and superiority."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gao-etal-2025-towards">
<titleInfo>
<title>Towards A Better Initial Policy Model For Scalable Long-CoT Reinforcement Learning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Bofei</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yejie</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yibo</namePart>
<namePart type="family">Miao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruoyu</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feifan</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Longhui</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianyu</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Baobao</namePart>
<namePart type="family">Chang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: ACL 2025</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wanxiang</namePart>
<namePart type="family">Che</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joyce</namePart>
<namePart type="family">Nabende</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Shutova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="given">Taher</namePart>
<namePart type="family">Pilehvar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Vienna, Austria</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-256-5</identifier>
</relatedItem>
<abstract>Long-CoT reasoning combined with reinforcement learning for large language models demonstrates remarkable performance and scalability. However, we observe that the initial policy model could significantly influence the final performance as well as the token efficiency. Additionally, there is a lack of systematic guidelines for obtaining a better initial policy model. To bridge this gap, we initiate a comprehensive investigation by activating the initial model using a variety of datasets with different data volumes and reasoning patterns. Then, we conduct a thorough analysis and comparison of the RL process for different initial models from the perspectives of upper bounds, diversity, and token efficiency, providing a deeper understanding and insight into the long-CoT RL. Based on our empirical results, we propose a systematic guideline and a novel Re-RFT method for constructing a better RL start point. Our experiment results based on the 14B model surpass the DeepSeek-R1-Distill-Qwen-14B by an average of 4.6%, demonstrating our approach’s effectiveness and superiority.</abstract>
<identifier type="citekey">gao-etal-2025-towards</identifier>
<identifier type="doi">10.18653/v1/2025.findings-acl.397</identifier>
<location>
<url>https://aclanthology.org/2025.findings-acl.397/</url>
</location>
<part>
<date>2025-07</date>
<extent unit="page">
<start>7652</start>
<end>7665</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Towards A Better Initial Policy Model For Scalable Long-CoT Reinforcement Learning
%A Gao, Bofei
%A Wang, Yejie
%A Miao, Yibo
%A Wu, Ruoyu
%A Song, Feifan
%A Yu, Longhui
%A Liu, Tianyu
%A Chang, Baobao
%Y Che, Wanxiang
%Y Nabende, Joyce
%Y Shutova, Ekaterina
%Y Pilehvar, Mohammad Taher
%S Findings of the Association for Computational Linguistics: ACL 2025
%D 2025
%8 July
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-256-5
%F gao-etal-2025-towards
%X Long-CoT reasoning combined with reinforcement learning for large language models demonstrates remarkable performance and scalability. However, we observe that the initial policy model could significantly influence the final performance as well as the token efficiency. Additionally, there is a lack of systematic guidelines for obtaining a better initial policy model. To bridge this gap, we initiate a comprehensive investigation by activating the initial model using a variety of datasets with different data volumes and reasoning patterns. Then, we conduct a thorough analysis and comparison of the RL process for different initial models from the perspectives of upper bounds, diversity, and token efficiency, providing a deeper understanding and insight into the long-CoT RL. Based on our empirical results, we propose a systematic guideline and a novel Re-RFT method for constructing a better RL start point. Our experiment results based on the 14B model surpass the DeepSeek-R1-Distill-Qwen-14B by an average of 4.6%, demonstrating our approach’s effectiveness and superiority.
%R 10.18653/v1/2025.findings-acl.397
%U https://aclanthology.org/2025.findings-acl.397/
%U https://doi.org/10.18653/v1/2025.findings-acl.397
%P 7652-7665
Markdown (Informal)
[Towards A Better Initial Policy Model For Scalable Long-CoT Reinforcement Learning](https://aclanthology.org/2025.findings-acl.397/) (Gao et al., Findings 2025)
ACL
- Bofei Gao, Yejie Wang, Yibo Miao, Ruoyu Wu, Feifan Song, Longhui Yu, Tianyu Liu, and Baobao Chang. 2025. Towards A Better Initial Policy Model For Scalable Long-CoT Reinforcement Learning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 7652–7665, Vienna, Austria. Association for Computational Linguistics.