@inproceedings{duan-etal-2024-negating,
title = "Negating Negatives: Alignment with Human Negative Samples via Distributional Dispreference Optimization",
author = "Duan, Shitong and
Yi, Xiaoyuan and
Zhang, Peng and
Liu, Yan and
Liu, Zheng and
Lu, Tun and
Xie, Xing and
Gu, Ning",
editor = "Al-Onaizan, Yaser and
Bansal, Mohit and
Chen, Yun-Nung",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
month = nov,
year = "2024",
address = "Miami, Florida, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.findings-emnlp.56/",
doi = "10.18653/v1/2024.findings-emnlp.56",
pages = "1012--1042",
abstract = "Large language models (LLMs) have revolutionized the role of AI, yet pose potential social risks. To steer LLMs towards human preference, alignment technologies have been introduced and gained increasing attention. Nevertheless, existing methods heavily rely on high-quality positive-negative training pairs, suffering from noisy positive responses that are barely distinguishable from negative ones. Given recent LLMs' proficiency in generating helpful responses, this work pivots towards a new research question: **can we achieve alignment using solely human-annotated negative samples, preserving helpfulness while reducing harmfulness?** For this purpose, we propose Distributional Dispreference Optimization (D$^2$O), which maximizes the discrepancy between dispreferred responses and the generated non-negative ones. In this way, D$^2$O effectively eschews harmful information without incorporating noisy positive samples, while avoiding collapse using self-generated responses as anchors. We demonstrate that D$^2$O can be regarded as learning a distributional preference model reflecting human dispreference against negative responses, which is theoretically an upper bound of the instance-level DPO. Extensive experiments manifest that our method achieves comparable generation quality and surpasses the latest strong baselines in producing less harmful and more informative responses with better training stability and faster convergence."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="duan-etal-2024-negating">
<titleInfo>
<title>Negating Negatives: Alignment with Human Negative Samples via Distributional Dispreference Optimization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shitong</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyuan</namePart>
<namePart type="family">Yi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yan</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tun</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xing</namePart>
<namePart type="family">Xie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ning</namePart>
<namePart type="family">Gu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2024</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yaser</namePart>
<namePart type="family">Al-Onaizan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohit</namePart>
<namePart type="family">Bansal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yun-Nung</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Miami, Florida, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Large language models (LLMs) have revolutionized the role of AI, yet pose potential social risks. To steer LLMs towards human preference, alignment technologies have been introduced and gained increasing attention. Nevertheless, existing methods heavily rely on high-quality positive-negative training pairs, suffering from noisy positive responses that are barely distinguishable from negative ones. Given recent LLMs’ proficiency in generating helpful responses, this work pivots towards a new research question: **can we achieve alignment using solely human-annotated negative samples, preserving helpfulness while reducing harmfulness?** For this purpose, we propose Distributional Dispreference Optimization (D²O), which maximizes the discrepancy between dispreferred responses and the generated non-negative ones. In this way, D²O effectively eschews harmful information without incorporating noisy positive samples, while avoiding collapse using self-generated responses as anchors. We demonstrate that D²O can be regarded as learning a distributional preference model reflecting human dispreference against negative responses, which is theoretically an upper bound of the instance-level DPO. Extensive experiments manifest that our method achieves comparable generation quality and surpasses the latest strong baselines in producing less harmful and more informative responses with better training stability and faster convergence.</abstract>
<identifier type="citekey">duan-etal-2024-negating</identifier>
<identifier type="doi">10.18653/v1/2024.findings-emnlp.56</identifier>
<location>
<url>https://aclanthology.org/2024.findings-emnlp.56/</url>
</location>
<part>
<date>2024-11</date>
<extent unit="page">
<start>1012</start>
<end>1042</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Negating Negatives: Alignment with Human Negative Samples via Distributional Dispreference Optimization
%A Duan, Shitong
%A Yi, Xiaoyuan
%A Zhang, Peng
%A Liu, Yan
%A Liu, Zheng
%A Lu, Tun
%A Xie, Xing
%A Gu, Ning
%Y Al-Onaizan, Yaser
%Y Bansal, Mohit
%Y Chen, Yun-Nung
%S Findings of the Association for Computational Linguistics: EMNLP 2024
%D 2024
%8 November
%I Association for Computational Linguistics
%C Miami, Florida, USA
%F duan-etal-2024-negating
%X Large language models (LLMs) have revolutionized the role of AI, yet pose potential social risks. To steer LLMs towards human preference, alignment technologies have been introduced and gained increasing attention. Nevertheless, existing methods heavily rely on high-quality positive-negative training pairs, suffering from noisy positive responses that are barely distinguishable from negative ones. Given recent LLMs’ proficiency in generating helpful responses, this work pivots towards a new research question: **can we achieve alignment using solely human-annotated negative samples, preserving helpfulness while reducing harmfulness?** For this purpose, we propose Distributional Dispreference Optimization (D²O), which maximizes the discrepancy between dispreferred responses and the generated non-negative ones. In this way, D²O effectively eschews harmful information without incorporating noisy positive samples, while avoiding collapse using self-generated responses as anchors. We demonstrate that D²O can be regarded as learning a distributional preference model reflecting human dispreference against negative responses, which is theoretically an upper bound of the instance-level DPO. Extensive experiments manifest that our method achieves comparable generation quality and surpasses the latest strong baselines in producing less harmful and more informative responses with better training stability and faster convergence.
%R 10.18653/v1/2024.findings-emnlp.56
%U https://aclanthology.org/2024.findings-emnlp.56/
%U https://doi.org/10.18653/v1/2024.findings-emnlp.56
%P 1012-1042
Markdown (Informal)
[Negating Negatives: Alignment with Human Negative Samples via Distributional Dispreference Optimization](https://aclanthology.org/2024.findings-emnlp.56/) (Duan et al., Findings 2024)
ACL
- Shitong Duan, Xiaoyuan Yi, Peng Zhang, Yan Liu, Zheng Liu, Tun Lu, Xing Xie, and Ning Gu. 2024. Negating Negatives: Alignment with Human Negative Samples via Distributional Dispreference Optimization. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1012–1042, Miami, Florida, USA. Association for Computational Linguistics.