<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hang-etal-2024-prior">
    <titleInfo>
        <title>Prior Constraints-based Reward Model Training for Aligning Large Language Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Zhou</namePart>
        <namePart type="family">Hang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Wang</namePart>
        <namePart type="family">Chenglong</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hu</namePart>
        <namePart type="family">Yimin</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Xiao</namePart>
        <namePart type="family">Tong</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhang</namePart>
        <namePart type="family">Chunliang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhu</namePart>
        <namePart type="family">Jingbo</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2024-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <language>
        <languageTerm type="text">eng</languageTerm>
    </language>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 23rd Chinese National Conference on Computational Linguistics (Volume 1: Main Conference)</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Maosong</namePart>
            <namePart type="family">Sun</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jiye</namePart>
            <namePart type="family">Liang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Xianpei</namePart>
            <namePart type="family">Han</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Zhiyuan</namePart>
            <namePart type="family">Liu</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Yulan</namePart>
            <namePart type="family">He</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Chinese Information Processing Society of China</publisher>
            <place>
                <placeTerm type="text">Taiyuan, China</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
    </relatedItem>
    <abstract>“Reinforcement learning with human feedback for aligning large language models (LLMs) trainsa reward model typically using ranking loss with comparison pairs. However, the training pro-cedure suffers from an inherent problem: the uncontrolled scaling of reward scores during rein-forcement learning due to the lack of constraints while training the reward model. This paperproposes a Prior Constraints-based Reward Model (PCRM) training method to mitigate thisproblem. PCRM incorporates prior constraints—specifically, length ratio and cosine similaritybetween outputs of each comparison pair—during reward model training to regulate optimiza-tion magnitude and control score margins. We comprehensively evaluate PCRM by examining itsrank correlation with human preferences and its effectiveness in aligning LLMs via RL. Exper-imental results demonstrate that PCRM significantly improves alignment performance by effec-tively constraining reward score scaling. As another bonus, our method is easily integrated intoarbitrary rank-based alignment methods, such as direct preference optimization, and can yieldconsistent improvement. The code is available at https://github.com/wangclnlp/DeepSpeed-Chat-Extension/tree/PCRM.”</abstract>
    <identifier type="citekey">hang-etal-2024-prior</identifier>
    <location>
        <url>https://aclanthology.org/2024.ccl-1.107/</url>
    </location>
    <part>
        <date>2024-07</date>
        <extent unit="page">
            <start>1395</start>
            <end>1407</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Prior Constraints-based Reward Model Training for Aligning Large Language Models
%A Hang, Zhou
%A Chenglong, Wang
%A Yimin, Hu
%A Tong, Xiao
%A Chunliang, Zhang
%A Jingbo, Zhu
%Y Sun, Maosong
%Y Liang, Jiye
%Y Han, Xianpei
%Y Liu, Zhiyuan
%Y He, Yulan
%S Proceedings of the 23rd Chinese National Conference on Computational Linguistics (Volume 1: Main Conference)
%D 2024
%8 July
%I Chinese Information Processing Society of China
%C Taiyuan, China
%G eng
%F hang-etal-2024-prior
%X “Reinforcement learning with human feedback for aligning large language models (LLMs) trainsa reward model typically using ranking loss with comparison pairs. However, the training pro-cedure suffers from an inherent problem: the uncontrolled scaling of reward scores during rein-forcement learning due to the lack of constraints while training the reward model. This paperproposes a Prior Constraints-based Reward Model (PCRM) training method to mitigate thisproblem. PCRM incorporates prior constraints—specifically, length ratio and cosine similaritybetween outputs of each comparison pair—during reward model training to regulate optimiza-tion magnitude and control score margins. We comprehensively evaluate PCRM by examining itsrank correlation with human preferences and its effectiveness in aligning LLMs via RL. Exper-imental results demonstrate that PCRM significantly improves alignment performance by effec-tively constraining reward score scaling. As another bonus, our method is easily integrated intoarbitrary rank-based alignment methods, such as direct preference optimization, and can yieldconsistent improvement. The code is available at https://github.com/wangclnlp/DeepSpeed-Chat-Extension/tree/PCRM.”
%U https://aclanthology.org/2024.ccl-1.107/
%P 1395-1407

Markdown (Informal)

[Prior Constraints-based Reward Model Training for Aligning Large Language Models](https://aclanthology.org/2024.ccl-1.107/) (Hang et al., CCL 2024)

• Prior Constraints-based Reward Model Training for Aligning Large Language Models (Hang et al., CCL 2024)

ACL

• Zhou Hang, Wang Chenglong, Hu Yimin, Xiao Tong, Zhang Chunliang, and Zhu Jingbo. 2024. Prior Constraints-based Reward Model Training for Aligning Large Language Models. In Proceedings of the 23rd Chinese National Conference on Computational Linguistics (Volume 1: Main Conference), pages 1395–1407, Taiyuan, China. Chinese Information Processing Society of China.