@inproceedings{jinnai-etal-2025-regularized,
title = "Regularized Best-of-N Sampling with Minimum {B}ayes Risk Objective for Language Model Alignment",
author = "Jinnai, Yuu and
Morimura, Tetsuro and
Ariu, Kaito and
Abe, Kenshi",
editor = "Chiruzzo, Luis and
Ritter, Alan and
Wang, Lu",
booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
month = apr,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.naacl-long.472/",
doi = "10.18653/v1/2025.naacl-long.472",
pages = "9321--9347",
ISBN = "979-8-89176-189-6",
abstract = "Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking when the accuracy of the reward model is not high enough. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. A common solution to prevent reward hacking in preference learning techniques is to optimize a reward using proximity regularization, which ensures that the language model remains close to the reference model. In this research, we propose MBR-BoN, a variant of BoN that aims to mitigate reward hacking at inference time by incorporating the Minimum Bayes Risk (MBR) objective as a proximity regularization term. We show empirically and analytically that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer for BoN sampling. We evaluate MBR-BoN on the AlpacaFarm and Anthropic{'}s hh-rlhf datasets and show that it outperforms both BoN sampling and MBR decoding. As an application of MBR-BoN, we use it to generate a pairwise preference learning dataset. Experimental results show that DPO models trained on a dataset generated with MBR-BoN outperform a DPO model generated with vanilla BoN."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jinnai-etal-2025-regularized">
<titleInfo>
<title>Regularized Best-of-N Sampling with Minimum Bayes Risk Objective for Language Model Alignment</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yuu</namePart>
<namePart type="family">Jinnai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tetsuro</namePart>
<namePart type="family">Morimura</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaito</namePart>
<namePart type="family">Ariu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kenshi</namePart>
<namePart type="family">Abe</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-04</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Luis</namePart>
<namePart type="family">Chiruzzo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alan</namePart>
<namePart type="family">Ritter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-189-6</identifier>
</relatedItem>
<abstract>Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking when the accuracy of the reward model is not high enough. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. A common solution to prevent reward hacking in preference learning techniques is to optimize a reward using proximity regularization, which ensures that the language model remains close to the reference model. In this research, we propose MBR-BoN, a variant of BoN that aims to mitigate reward hacking at inference time by incorporating the Minimum Bayes Risk (MBR) objective as a proximity regularization term. We show empirically and analytically that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer for BoN sampling. We evaluate MBR-BoN on the AlpacaFarm and Anthropic’s hh-rlhf datasets and show that it outperforms both BoN sampling and MBR decoding. As an application of MBR-BoN, we use it to generate a pairwise preference learning dataset. Experimental results show that DPO models trained on a dataset generated with MBR-BoN outperform a DPO model generated with vanilla BoN.</abstract>
<identifier type="citekey">jinnai-etal-2025-regularized</identifier>
<identifier type="doi">10.18653/v1/2025.naacl-long.472</identifier>
<location>
<url>https://aclanthology.org/2025.naacl-long.472/</url>
</location>
<part>
<date>2025-04</date>
<extent unit="page">
<start>9321</start>
<end>9347</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Regularized Best-of-N Sampling with Minimum Bayes Risk Objective for Language Model Alignment
%A Jinnai, Yuu
%A Morimura, Tetsuro
%A Ariu, Kaito
%A Abe, Kenshi
%Y Chiruzzo, Luis
%Y Ritter, Alan
%Y Wang, Lu
%S Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)
%D 2025
%8 April
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-189-6
%F jinnai-etal-2025-regularized
%X Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking when the accuracy of the reward model is not high enough. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. A common solution to prevent reward hacking in preference learning techniques is to optimize a reward using proximity regularization, which ensures that the language model remains close to the reference model. In this research, we propose MBR-BoN, a variant of BoN that aims to mitigate reward hacking at inference time by incorporating the Minimum Bayes Risk (MBR) objective as a proximity regularization term. We show empirically and analytically that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer for BoN sampling. We evaluate MBR-BoN on the AlpacaFarm and Anthropic’s hh-rlhf datasets and show that it outperforms both BoN sampling and MBR decoding. As an application of MBR-BoN, we use it to generate a pairwise preference learning dataset. Experimental results show that DPO models trained on a dataset generated with MBR-BoN outperform a DPO model generated with vanilla BoN.
%R 10.18653/v1/2025.naacl-long.472
%U https://aclanthology.org/2025.naacl-long.472/
%U https://doi.org/10.18653/v1/2025.naacl-long.472
%P 9321-9347
Markdown (Informal)
[Regularized Best-of-N Sampling with Minimum Bayes Risk Objective for Language Model Alignment](https://aclanthology.org/2025.naacl-long.472/) (Jinnai et al., NAACL 2025)
ACL