@inproceedings{mcdougall-etal-2024-copy,
title = "Copy Suppression: Comprehensively Understanding a Motif in Language Model Attention Heads",
author = "McDougall, Callum Stuart and
Conmy, Arthur and
Rushing, Cody and
McGrath, Thomas and
Nanda, Neel",
editor = "Belinkov, Yonatan and
Kim, Najoung and
Jumelet, Jaap and
Mohebbi, Hosein and
Mueller, Aaron and
Chen, Hanjie",
booktitle = "Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP",
month = nov,
year = "2024",
address = "Miami, Florida, US",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.blackboxnlp-1.22",
pages = "337--363",
abstract = "We present the copy suppression motif: an algorithm implemented by attention heads in large language models that reduces loss.If i) language model components in earlier layers predict a certain token, ii) this token appears earlier in the context and iii) later attention heads in the model suppress prediction of the token, then this is copy suppression. To show the importance of copy suppression, we focus on reverse-engineering attention head 10.7 (L10H7) in GPT-2 Small. This head suppresses naive copying behavior which improves overall model calibration, which explains why multiple prior works studying certain narrow tasks found negative heads that systematically favored the wrong answer. We uncover the mechanism that the negative heads use for copy suppression with weights-based evidence and are able to explain 76.9{\%} of the impact of L10H7 in GPT-2 Small, by this motif alone.To the best of our knowledge, this is the most comprehensive description of the complete role of a component in a language model to date. One major effect of copy suppression is its role in self-repair. Self-repair refers to how ablating crucial model components results in downstream neural network parts compensating for this ablation. Copy suppression leads to self-repair: if an initial overconfident copier is ablated, then there is nothing to suppress. We show that self-repair is implemented by several mechanisms, one of which is copy suppression, which explains 39{\%} of the behavior in a narrow task. Interactive visualizations of the copy suppression phenomena may be seen at our web app https://copy-suppression.streamlit.app/.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="mcdougall-etal-2024-copy">
<titleInfo>
<title>Copy Suppression: Comprehensively Understanding a Motif in Language Model Attention Heads</title>
</titleInfo>
<name type="personal">
<namePart type="given">Callum</namePart>
<namePart type="given">Stuart</namePart>
<namePart type="family">McDougall</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arthur</namePart>
<namePart type="family">Conmy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cody</namePart>
<namePart type="family">Rushing</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thomas</namePart>
<namePart type="family">McGrath</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neel</namePart>
<namePart type="family">Nanda</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>Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yonatan</namePart>
<namePart type="family">Belinkov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Najoung</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jaap</namePart>
<namePart type="family">Jumelet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hosein</namePart>
<namePart type="family">Mohebbi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aaron</namePart>
<namePart type="family">Mueller</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanjie</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, US</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>We present the copy suppression motif: an algorithm implemented by attention heads in large language models that reduces loss.If i) language model components in earlier layers predict a certain token, ii) this token appears earlier in the context and iii) later attention heads in the model suppress prediction of the token, then this is copy suppression. To show the importance of copy suppression, we focus on reverse-engineering attention head 10.7 (L10H7) in GPT-2 Small. This head suppresses naive copying behavior which improves overall model calibration, which explains why multiple prior works studying certain narrow tasks found negative heads that systematically favored the wrong answer. We uncover the mechanism that the negative heads use for copy suppression with weights-based evidence and are able to explain 76.9% of the impact of L10H7 in GPT-2 Small, by this motif alone.To the best of our knowledge, this is the most comprehensive description of the complete role of a component in a language model to date. One major effect of copy suppression is its role in self-repair. Self-repair refers to how ablating crucial model components results in downstream neural network parts compensating for this ablation. Copy suppression leads to self-repair: if an initial overconfident copier is ablated, then there is nothing to suppress. We show that self-repair is implemented by several mechanisms, one of which is copy suppression, which explains 39% of the behavior in a narrow task. Interactive visualizations of the copy suppression phenomena may be seen at our web app https://copy-suppression.streamlit.app/.</abstract>
<identifier type="citekey">mcdougall-etal-2024-copy</identifier>
<location>
<url>https://aclanthology.org/2024.blackboxnlp-1.22</url>
</location>
<part>
<date>2024-11</date>
<extent unit="page">
<start>337</start>
<end>363</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Copy Suppression: Comprehensively Understanding a Motif in Language Model Attention Heads
%A McDougall, Callum Stuart
%A Conmy, Arthur
%A Rushing, Cody
%A McGrath, Thomas
%A Nanda, Neel
%Y Belinkov, Yonatan
%Y Kim, Najoung
%Y Jumelet, Jaap
%Y Mohebbi, Hosein
%Y Mueller, Aaron
%Y Chen, Hanjie
%S Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP
%D 2024
%8 November
%I Association for Computational Linguistics
%C Miami, Florida, US
%F mcdougall-etal-2024-copy
%X We present the copy suppression motif: an algorithm implemented by attention heads in large language models that reduces loss.If i) language model components in earlier layers predict a certain token, ii) this token appears earlier in the context and iii) later attention heads in the model suppress prediction of the token, then this is copy suppression. To show the importance of copy suppression, we focus on reverse-engineering attention head 10.7 (L10H7) in GPT-2 Small. This head suppresses naive copying behavior which improves overall model calibration, which explains why multiple prior works studying certain narrow tasks found negative heads that systematically favored the wrong answer. We uncover the mechanism that the negative heads use for copy suppression with weights-based evidence and are able to explain 76.9% of the impact of L10H7 in GPT-2 Small, by this motif alone.To the best of our knowledge, this is the most comprehensive description of the complete role of a component in a language model to date. One major effect of copy suppression is its role in self-repair. Self-repair refers to how ablating crucial model components results in downstream neural network parts compensating for this ablation. Copy suppression leads to self-repair: if an initial overconfident copier is ablated, then there is nothing to suppress. We show that self-repair is implemented by several mechanisms, one of which is copy suppression, which explains 39% of the behavior in a narrow task. Interactive visualizations of the copy suppression phenomena may be seen at our web app https://copy-suppression.streamlit.app/.
%U https://aclanthology.org/2024.blackboxnlp-1.22
%P 337-363
Markdown (Informal)
[Copy Suppression: Comprehensively Understanding a Motif in Language Model Attention Heads](https://aclanthology.org/2024.blackboxnlp-1.22) (McDougall et al., BlackboxNLP 2024)
ACL