@inproceedings{ahrabian-etal-2025-practical,
title = "A Practical Analysis of Human Alignment with *{PO}",
author = "Ahrabian, Kian and
Lin, Xihui and
Patra, Barun and
Chaudhary, Vishrav and
Benhaim, Alon and
Pujara, Jay and
Song, Xia",
editor = "Chiruzzo, Luis and
Ritter, Alan and
Wang, Lu",
booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
month = apr,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-naacl.446/",
doi = "10.18653/v1/2025.findings-naacl.446",
pages = "8013--8021",
ISBN = "979-8-89176-195-7",
abstract = "At the forefront of state-of-the-art human alignment methods are preference optimization methods (*PO). Prior research has often concentrated on identifying the best-performing method, typically involving a grid search over hyperparameters, which can be impractical for general practitioners. In this paper, we examine the robustness of existing state-of-the-art methods to varying hyperparameters in a realistic out-of-distribution (OOD) scenario that mirrors real-world applications of human alignment. Our goal is to empirically find the method that increases the likelihood of achieving better results through the lens of various metrics, such as KL divergence and response length. We also introduce LN-DPO, a simple length-normalized version of DPO that is more stable across hyperparameters, effectively reduces the average response length, and improves performance. Our analysis of state-of-the-art reference-free (i.e., SimPO) and reference-dependent (i.e., DPO and LN-DPO) methods reveals that they perform similarly at their peak (i.e., best possible scenario). However, we uncover that the pattern of change in performance greatly varies as we move away from the best possible scenario."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ahrabian-etal-2025-practical">
<titleInfo>
<title>A Practical Analysis of Human Alignment with *PO</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kian</namePart>
<namePart type="family">Ahrabian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xihui</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barun</namePart>
<namePart type="family">Patra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vishrav</namePart>
<namePart type="family">Chaudhary</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alon</namePart>
<namePart type="family">Benhaim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jay</namePart>
<namePart type="family">Pujara</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xia</namePart>
<namePart type="family">Song</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>Findings of the Association for Computational Linguistics: NAACL 2025</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">979-8-89176-195-7</identifier>
</relatedItem>
<abstract>At the forefront of state-of-the-art human alignment methods are preference optimization methods (*PO). Prior research has often concentrated on identifying the best-performing method, typically involving a grid search over hyperparameters, which can be impractical for general practitioners. In this paper, we examine the robustness of existing state-of-the-art methods to varying hyperparameters in a realistic out-of-distribution (OOD) scenario that mirrors real-world applications of human alignment. Our goal is to empirically find the method that increases the likelihood of achieving better results through the lens of various metrics, such as KL divergence and response length. We also introduce LN-DPO, a simple length-normalized version of DPO that is more stable across hyperparameters, effectively reduces the average response length, and improves performance. Our analysis of state-of-the-art reference-free (i.e., SimPO) and reference-dependent (i.e., DPO and LN-DPO) methods reveals that they perform similarly at their peak (i.e., best possible scenario). However, we uncover that the pattern of change in performance greatly varies as we move away from the best possible scenario.</abstract>
<identifier type="citekey">ahrabian-etal-2025-practical</identifier>
<identifier type="doi">10.18653/v1/2025.findings-naacl.446</identifier>
<location>
<url>https://aclanthology.org/2025.findings-naacl.446/</url>
</location>
<part>
<date>2025-04</date>
<extent unit="page">
<start>8013</start>
<end>8021</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T A Practical Analysis of Human Alignment with *PO
%A Ahrabian, Kian
%A Lin, Xihui
%A Patra, Barun
%A Chaudhary, Vishrav
%A Benhaim, Alon
%A Pujara, Jay
%A Song, Xia
%Y Chiruzzo, Luis
%Y Ritter, Alan
%Y Wang, Lu
%S Findings of the Association for Computational Linguistics: NAACL 2025
%D 2025
%8 April
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-195-7
%F ahrabian-etal-2025-practical
%X At the forefront of state-of-the-art human alignment methods are preference optimization methods (*PO). Prior research has often concentrated on identifying the best-performing method, typically involving a grid search over hyperparameters, which can be impractical for general practitioners. In this paper, we examine the robustness of existing state-of-the-art methods to varying hyperparameters in a realistic out-of-distribution (OOD) scenario that mirrors real-world applications of human alignment. Our goal is to empirically find the method that increases the likelihood of achieving better results through the lens of various metrics, such as KL divergence and response length. We also introduce LN-DPO, a simple length-normalized version of DPO that is more stable across hyperparameters, effectively reduces the average response length, and improves performance. Our analysis of state-of-the-art reference-free (i.e., SimPO) and reference-dependent (i.e., DPO and LN-DPO) methods reveals that they perform similarly at their peak (i.e., best possible scenario). However, we uncover that the pattern of change in performance greatly varies as we move away from the best possible scenario.
%R 10.18653/v1/2025.findings-naacl.446
%U https://aclanthology.org/2025.findings-naacl.446/
%U https://doi.org/10.18653/v1/2025.findings-naacl.446
%P 8013-8021
Markdown (Informal)
[A Practical Analysis of Human Alignment with *PO](https://aclanthology.org/2025.findings-naacl.446/) (Ahrabian et al., Findings 2025)
ACL
- Kian Ahrabian, Xihui Lin, Barun Patra, Vishrav Chaudhary, Alon Benhaim, Jay Pujara, and Xia Song. 2025. A Practical Analysis of Human Alignment with *PO. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 8013–8021, Albuquerque, New Mexico. Association for Computational Linguistics.