@inproceedings{liu-etal-2024-good-llms,
title = "How Good Are {LLM}s at Out-of-Distribution Detection?",
author = "Liu, Bo and
Zhan, Li-Ming and
Lu, Zexin and
Feng, Yujie and
Xue, Lei and
Wu, Xiao-Ming",
editor = "Calzolari, Nicoletta and
Kan, Min-Yen and
Hoste, Veronique and
Lenci, Alessandro and
Sakti, Sakriani and
Xue, Nianwen",
booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
month = may,
year = "2024",
address = "Torino, Italia",
publisher = "ELRA and ICCL",
url = "https://aclanthology.org/2024.lrec-main.720",
pages = "8211--8222",
abstract = "Out-of-distribution (OOD) detection plays a vital role in enhancing the reliability of machine learning models. As large language models (LLMs) become more prevalent, the applicability of prior research on OOD detection that utilized smaller-scale Transformers such as BERT, RoBERTa, and GPT-2 may be challenged, due to the significant differences in the scale of these models, their pre-training objectives, and the paradigms used for inference. This paper initiates a pioneering empirical investigation into the OOD detection capabilities of LLMs, focusing on the LLaMA series ranging from 7B to 65B in size. We thoroughly evaluate commonly used OOD detectors, examining their performance in both zero-grad and fine-tuning scenarios. Notably, we alter previous discriminative in-distribution fine-tuning into generative fine-tuning, aligning the pre-training objective of LLMs with downstream tasks. Our findings unveil that a simple cosine distance OOD detector demonstrates superior efficacy, outperforming other OOD detectors. We provide an intriguing explanation for this phenomenon by highlighting the isotropic nature of the embedding spaces of LLMs, which distinctly contrasts with the anisotropic property observed in smaller BERT family models. The new insight enhances our understanding of how LLMs detect OOD data, thereby enhancing their adaptability and reliability in dynamic environments. We have released the source code at \url{https://github.com/Awenbocc/LLM-OOD} for other researchers to reproduce our results.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2024-good-llms">
<titleInfo>
<title>How Good Are LLMs at Out-of-Distribution Detection?</title>
</titleInfo>
<name type="personal">
<namePart type="given">Bo</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Li-Ming</namePart>
<namePart type="family">Zhan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zexin</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yujie</namePart>
<namePart type="family">Feng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lei</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiao-Ming</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicoletta</namePart>
<namePart type="family">Calzolari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Min-Yen</namePart>
<namePart type="family">Kan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Veronique</namePart>
<namePart type="family">Hoste</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">Lenci</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakriani</namePart>
<namePart type="family">Sakti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nianwen</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA and ICCL</publisher>
<place>
<placeTerm type="text">Torino, Italia</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Out-of-distribution (OOD) detection plays a vital role in enhancing the reliability of machine learning models. As large language models (LLMs) become more prevalent, the applicability of prior research on OOD detection that utilized smaller-scale Transformers such as BERT, RoBERTa, and GPT-2 may be challenged, due to the significant differences in the scale of these models, their pre-training objectives, and the paradigms used for inference. This paper initiates a pioneering empirical investigation into the OOD detection capabilities of LLMs, focusing on the LLaMA series ranging from 7B to 65B in size. We thoroughly evaluate commonly used OOD detectors, examining their performance in both zero-grad and fine-tuning scenarios. Notably, we alter previous discriminative in-distribution fine-tuning into generative fine-tuning, aligning the pre-training objective of LLMs with downstream tasks. Our findings unveil that a simple cosine distance OOD detector demonstrates superior efficacy, outperforming other OOD detectors. We provide an intriguing explanation for this phenomenon by highlighting the isotropic nature of the embedding spaces of LLMs, which distinctly contrasts with the anisotropic property observed in smaller BERT family models. The new insight enhances our understanding of how LLMs detect OOD data, thereby enhancing their adaptability and reliability in dynamic environments. We have released the source code at https://github.com/Awenbocc/LLM-OOD for other researchers to reproduce our results.</abstract>
<identifier type="citekey">liu-etal-2024-good-llms</identifier>
<location>
<url>https://aclanthology.org/2024.lrec-main.720</url>
</location>
<part>
<date>2024-05</date>
<extent unit="page">
<start>8211</start>
<end>8222</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Good Are LLMs at Out-of-Distribution Detection?
%A Liu, Bo
%A Zhan, Li-Ming
%A Lu, Zexin
%A Feng, Yujie
%A Xue, Lei
%A Wu, Xiao-Ming
%Y Calzolari, Nicoletta
%Y Kan, Min-Yen
%Y Hoste, Veronique
%Y Lenci, Alessandro
%Y Sakti, Sakriani
%Y Xue, Nianwen
%S Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)
%D 2024
%8 May
%I ELRA and ICCL
%C Torino, Italia
%F liu-etal-2024-good-llms
%X Out-of-distribution (OOD) detection plays a vital role in enhancing the reliability of machine learning models. As large language models (LLMs) become more prevalent, the applicability of prior research on OOD detection that utilized smaller-scale Transformers such as BERT, RoBERTa, and GPT-2 may be challenged, due to the significant differences in the scale of these models, their pre-training objectives, and the paradigms used for inference. This paper initiates a pioneering empirical investigation into the OOD detection capabilities of LLMs, focusing on the LLaMA series ranging from 7B to 65B in size. We thoroughly evaluate commonly used OOD detectors, examining their performance in both zero-grad and fine-tuning scenarios. Notably, we alter previous discriminative in-distribution fine-tuning into generative fine-tuning, aligning the pre-training objective of LLMs with downstream tasks. Our findings unveil that a simple cosine distance OOD detector demonstrates superior efficacy, outperforming other OOD detectors. We provide an intriguing explanation for this phenomenon by highlighting the isotropic nature of the embedding spaces of LLMs, which distinctly contrasts with the anisotropic property observed in smaller BERT family models. The new insight enhances our understanding of how LLMs detect OOD data, thereby enhancing their adaptability and reliability in dynamic environments. We have released the source code at https://github.com/Awenbocc/LLM-OOD for other researchers to reproduce our results.
%U https://aclanthology.org/2024.lrec-main.720
%P 8211-8222
Markdown (Informal)
[How Good Are LLMs at Out-of-Distribution Detection?](https://aclanthology.org/2024.lrec-main.720) (Liu et al., LREC-COLING 2024)
ACL
- Bo Liu, Li-Ming Zhan, Zexin Lu, Yujie Feng, Lei Xue, and Xiao-Ming Wu. 2024. How Good Are LLMs at Out-of-Distribution Detection?. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 8211–8222, Torino, Italia. ELRA and ICCL.