@inproceedings{devanathan-etal-2024-paradox,
title = "The Paradox of Preference: A Study on {LLM} Alignment Algorithms and Data Acquisition Methods",
author = "Devanathan, Rishikesh and
Nathan, Varun and
Kumar, Ayush",
editor = "Tafreshi, Shabnam and
Akula, Arjun and
Sedoc, Jo{\~a}o and
Drozd, Aleksandr and
Rogers, Anna and
Rumshisky, Anna",
booktitle = "Proceedings of the Fifth Workshop on Insights from Negative Results in NLP",
month = jun,
year = "2024",
address = "Mexico City, Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.insights-1.16",
doi = "10.18653/v1/2024.insights-1.16",
pages = "135--147",
abstract = "This research investigates the impact of preference annotation acquisition methods on the performance of LLM alignment algorithms, including Direct Preference Optimization (DPO), Identity Preference Optimization (IPO), and Conservative DPO (cDPO), compared to Supervised Fine-Tuning (SFT) in NLP tasks. We analyze the influence of LLM and human-based preferences on algorithm performance, considering data volume and quality. Additionally, we assess DPO{'}s vulnerability to overfitting and IPO{'}s resilience against it, addressing four main research questions. Using the GAIR dataset and Zephyr-7b as the SFT model, we reveal unexpected negative outcomes. Specifically, DPO trained on LLM preferences outperforms human preferences, contrary to expectations. Moreover, there{'}s no correlation between preference data volume or quality and algorithm performance. Contrary to expectations, DPO shows no overfitting in both human and LLM preference datasets. Surprisingly, cDPO doesn{'}t fare better than DPO under flip noise. Our findings highlight the complexities of preference annotation methods and underscore the importance of scrutinizing negative results in NLP algorithm research.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="devanathan-etal-2024-paradox">
<titleInfo>
<title>The Paradox of Preference: A Study on LLM Alignment Algorithms and Data Acquisition Methods</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rishikesh</namePart>
<namePart type="family">Devanathan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Varun</namePart>
<namePart type="family">Nathan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayush</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifth Workshop on Insights from Negative Results in NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shabnam</namePart>
<namePart type="family">Tafreshi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arjun</namePart>
<namePart type="family">Akula</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">João</namePart>
<namePart type="family">Sedoc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aleksandr</namePart>
<namePart type="family">Drozd</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Rogers</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Rumshisky</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Mexico City, Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>This research investigates the impact of preference annotation acquisition methods on the performance of LLM alignment algorithms, including Direct Preference Optimization (DPO), Identity Preference Optimization (IPO), and Conservative DPO (cDPO), compared to Supervised Fine-Tuning (SFT) in NLP tasks. We analyze the influence of LLM and human-based preferences on algorithm performance, considering data volume and quality. Additionally, we assess DPO’s vulnerability to overfitting and IPO’s resilience against it, addressing four main research questions. Using the GAIR dataset and Zephyr-7b as the SFT model, we reveal unexpected negative outcomes. Specifically, DPO trained on LLM preferences outperforms human preferences, contrary to expectations. Moreover, there’s no correlation between preference data volume or quality and algorithm performance. Contrary to expectations, DPO shows no overfitting in both human and LLM preference datasets. Surprisingly, cDPO doesn’t fare better than DPO under flip noise. Our findings highlight the complexities of preference annotation methods and underscore the importance of scrutinizing negative results in NLP algorithm research.</abstract>
<identifier type="citekey">devanathan-etal-2024-paradox</identifier>
<identifier type="doi">10.18653/v1/2024.insights-1.16</identifier>
<location>
<url>https://aclanthology.org/2024.insights-1.16</url>
</location>
<part>
<date>2024-06</date>
<extent unit="page">
<start>135</start>
<end>147</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Paradox of Preference: A Study on LLM Alignment Algorithms and Data Acquisition Methods
%A Devanathan, Rishikesh
%A Nathan, Varun
%A Kumar, Ayush
%Y Tafreshi, Shabnam
%Y Akula, Arjun
%Y Sedoc, João
%Y Drozd, Aleksandr
%Y Rogers, Anna
%Y Rumshisky, Anna
%S Proceedings of the Fifth Workshop on Insights from Negative Results in NLP
%D 2024
%8 June
%I Association for Computational Linguistics
%C Mexico City, Mexico
%F devanathan-etal-2024-paradox
%X This research investigates the impact of preference annotation acquisition methods on the performance of LLM alignment algorithms, including Direct Preference Optimization (DPO), Identity Preference Optimization (IPO), and Conservative DPO (cDPO), compared to Supervised Fine-Tuning (SFT) in NLP tasks. We analyze the influence of LLM and human-based preferences on algorithm performance, considering data volume and quality. Additionally, we assess DPO’s vulnerability to overfitting and IPO’s resilience against it, addressing four main research questions. Using the GAIR dataset and Zephyr-7b as the SFT model, we reveal unexpected negative outcomes. Specifically, DPO trained on LLM preferences outperforms human preferences, contrary to expectations. Moreover, there’s no correlation between preference data volume or quality and algorithm performance. Contrary to expectations, DPO shows no overfitting in both human and LLM preference datasets. Surprisingly, cDPO doesn’t fare better than DPO under flip noise. Our findings highlight the complexities of preference annotation methods and underscore the importance of scrutinizing negative results in NLP algorithm research.
%R 10.18653/v1/2024.insights-1.16
%U https://aclanthology.org/2024.insights-1.16
%U https://doi.org/10.18653/v1/2024.insights-1.16
%P 135-147
Markdown (Informal)
[The Paradox of Preference: A Study on LLM Alignment Algorithms and Data Acquisition Methods](https://aclanthology.org/2024.insights-1.16) (Devanathan et al., insights-WS 2024)
ACL