@inproceedings{tigges-etal-2024-llm,
title = "{LLM} Circuit Analyses Are Consistent Across Training and Scale",
author = "Tigges, Curt and
Hanna, Michael and
Yu, Qinan and
Biderman, Stella",
editor = "Zhao, Chen and
Mosbach, Marius and
Atanasova, Pepa and
Goldfarb-Tarrent, Seraphina and
Hase, Peter and
Hosseini, Arian and
Elbayad, Maha and
Pezzelle, Sandro and
Mozes, Maximilian",
booktitle = "Proceedings of the 9th Workshop on Representation Learning for NLP (RepL4NLP-2024)",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.repl4nlp-1.22",
pages = "290--303",
abstract = "Most currently deployed large language models (LLMs) undergo continuous training or additional finetuning. By contrast, most research into LLMs{'} internal mechanisms focuses on models at one snapshot in time (the end of pre-training), raising the question of whether their results generalize to real-world settings. Existing studies of mechanisms over time focus on encoder-only or toy models, which differ significantly from most deployed models. In this study, we track how model mechanisms, operationalized as circuits, emerge and evolve across 300 billion tokens of training in decoder-only LLMs, in models ranging from 70 million to 2.8 billion parameters. We find that task abilities and the functional components that support them emerge consistently at similar token counts across scale. Moreover, although such components may be implemented by different attention heads over time, the overarching algorithm that they implement remains. Surprisingly, both these algorithms and the types of components involved therein tend to replicate across model scale. Finally, we find that circuit size correlates with model size and can fluctuate considerably over time even when the same algorithm is implemented. These results suggest that circuit analyses conducted on small models at the end of pre-training can provide insights that still apply after additional training and over model scale.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tigges-etal-2024-llm">
<titleInfo>
<title>LLM Circuit Analyses Are Consistent Across Training and Scale</title>
</titleInfo>
<name type="personal">
<namePart type="given">Curt</namePart>
<namePart type="family">Tigges</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Hanna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qinan</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stella</namePart>
<namePart type="family">Biderman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 9th Workshop on Representation Learning for NLP (RepL4NLP-2024)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chen</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marius</namePart>
<namePart type="family">Mosbach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seraphina</namePart>
<namePart type="family">Goldfarb-Tarrent</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peter</namePart>
<namePart type="family">Hase</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arian</namePart>
<namePart type="family">Hosseini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maha</namePart>
<namePart type="family">Elbayad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sandro</namePart>
<namePart type="family">Pezzelle</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maximilian</namePart>
<namePart type="family">Mozes</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Bangkok, Thailand</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Most currently deployed large language models (LLMs) undergo continuous training or additional finetuning. By contrast, most research into LLMs’ internal mechanisms focuses on models at one snapshot in time (the end of pre-training), raising the question of whether their results generalize to real-world settings. Existing studies of mechanisms over time focus on encoder-only or toy models, which differ significantly from most deployed models. In this study, we track how model mechanisms, operationalized as circuits, emerge and evolve across 300 billion tokens of training in decoder-only LLMs, in models ranging from 70 million to 2.8 billion parameters. We find that task abilities and the functional components that support them emerge consistently at similar token counts across scale. Moreover, although such components may be implemented by different attention heads over time, the overarching algorithm that they implement remains. Surprisingly, both these algorithms and the types of components involved therein tend to replicate across model scale. Finally, we find that circuit size correlates with model size and can fluctuate considerably over time even when the same algorithm is implemented. These results suggest that circuit analyses conducted on small models at the end of pre-training can provide insights that still apply after additional training and over model scale.</abstract>
<identifier type="citekey">tigges-etal-2024-llm</identifier>
<location>
<url>https://aclanthology.org/2024.repl4nlp-1.22</url>
</location>
<part>
<date>2024-08</date>
<extent unit="page">
<start>290</start>
<end>303</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LLM Circuit Analyses Are Consistent Across Training and Scale
%A Tigges, Curt
%A Hanna, Michael
%A Yu, Qinan
%A Biderman, Stella
%Y Zhao, Chen
%Y Mosbach, Marius
%Y Atanasova, Pepa
%Y Goldfarb-Tarrent, Seraphina
%Y Hase, Peter
%Y Hosseini, Arian
%Y Elbayad, Maha
%Y Pezzelle, Sandro
%Y Mozes, Maximilian
%S Proceedings of the 9th Workshop on Representation Learning for NLP (RepL4NLP-2024)
%D 2024
%8 August
%I Association for Computational Linguistics
%C Bangkok, Thailand
%F tigges-etal-2024-llm
%X Most currently deployed large language models (LLMs) undergo continuous training or additional finetuning. By contrast, most research into LLMs’ internal mechanisms focuses on models at one snapshot in time (the end of pre-training), raising the question of whether their results generalize to real-world settings. Existing studies of mechanisms over time focus on encoder-only or toy models, which differ significantly from most deployed models. In this study, we track how model mechanisms, operationalized as circuits, emerge and evolve across 300 billion tokens of training in decoder-only LLMs, in models ranging from 70 million to 2.8 billion parameters. We find that task abilities and the functional components that support them emerge consistently at similar token counts across scale. Moreover, although such components may be implemented by different attention heads over time, the overarching algorithm that they implement remains. Surprisingly, both these algorithms and the types of components involved therein tend to replicate across model scale. Finally, we find that circuit size correlates with model size and can fluctuate considerably over time even when the same algorithm is implemented. These results suggest that circuit analyses conducted on small models at the end of pre-training can provide insights that still apply after additional training and over model scale.
%U https://aclanthology.org/2024.repl4nlp-1.22
%P 290-303
Markdown (Informal)
[LLM Circuit Analyses Are Consistent Across Training and Scale](https://aclanthology.org/2024.repl4nlp-1.22) (Tigges et al., RepL4NLP-WS 2024)
ACL