@inproceedings{hamilton-2024-detecting,
title = "Detecting Mode Collapse in Language Models via Narration",
author = "Hamilton, Sil",
editor = "Miceli-Barone, Antonio Valerio and
Barez, Fazl and
Cohen, Shay and
Voita, Elena and
Germann, Ulrich and
Lukasik, Michal",
booktitle = "Proceedings of the First edition of the Workshop on the Scaling Behavior of Large Language Models (SCALE-LLM 2024)",
month = mar,
year = "2024",
address = "St. Julian{'}s, Malta",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.scalellm-1.5",
pages = "65--72",
abstract = "No two authors write alike. Personal flourishes invoked in written narratives, from lexicon to rhetorical devices, imply a particular author{---}what literary theorists label the implied or virtual author; distinct from the real author or narrator of a text. Early large language models trained on unfiltered training sets drawn from a variety of discordant sources yielded incoherent personalities, problematic for conversational tasks but proving useful for sampling literature from multiple perspectives. Successes in alignment research in recent years have allowed researchers to impose subjectively consistent personae on language models via instruction tuning and reinforcement learning from human feedback (RLHF), but whether aligned models retain the ability to model an arbitrary virtual author has received little scrutiny. By studying 4,374 stories sampled from three OpenAI language models, we show successive versions of GPT-3 suffer from increasing degrees of {``}mode collapse{''} whereby overfitting the model during alignment constrains it from generalizing over authorship: models suffering from mode collapse become unable to assume a multiplicity of perspectives. Our method and results are significant for researchers seeking to employ language models in sociological simulations.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hamilton-2024-detecting">
<titleInfo>
<title>Detecting Mode Collapse in Language Models via Narration</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sil</namePart>
<namePart type="family">Hamilton</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-03</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the First edition of the Workshop on the Scaling Behavior of Large Language Models (SCALE-LLM 2024)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Antonio</namePart>
<namePart type="given">Valerio</namePart>
<namePart type="family">Miceli-Barone</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fazl</namePart>
<namePart type="family">Barez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shay</namePart>
<namePart type="family">Cohen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elena</namePart>
<namePart type="family">Voita</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ulrich</namePart>
<namePart type="family">Germann</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michal</namePart>
<namePart type="family">Lukasik</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">St. Julian’s, Malta</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>No two authors write alike. Personal flourishes invoked in written narratives, from lexicon to rhetorical devices, imply a particular author—what literary theorists label the implied or virtual author; distinct from the real author or narrator of a text. Early large language models trained on unfiltered training sets drawn from a variety of discordant sources yielded incoherent personalities, problematic for conversational tasks but proving useful for sampling literature from multiple perspectives. Successes in alignment research in recent years have allowed researchers to impose subjectively consistent personae on language models via instruction tuning and reinforcement learning from human feedback (RLHF), but whether aligned models retain the ability to model an arbitrary virtual author has received little scrutiny. By studying 4,374 stories sampled from three OpenAI language models, we show successive versions of GPT-3 suffer from increasing degrees of “mode collapse” whereby overfitting the model during alignment constrains it from generalizing over authorship: models suffering from mode collapse become unable to assume a multiplicity of perspectives. Our method and results are significant for researchers seeking to employ language models in sociological simulations.</abstract>
<identifier type="citekey">hamilton-2024-detecting</identifier>
<location>
<url>https://aclanthology.org/2024.scalellm-1.5</url>
</location>
<part>
<date>2024-03</date>
<extent unit="page">
<start>65</start>
<end>72</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Detecting Mode Collapse in Language Models via Narration
%A Hamilton, Sil
%Y Miceli-Barone, Antonio Valerio
%Y Barez, Fazl
%Y Cohen, Shay
%Y Voita, Elena
%Y Germann, Ulrich
%Y Lukasik, Michal
%S Proceedings of the First edition of the Workshop on the Scaling Behavior of Large Language Models (SCALE-LLM 2024)
%D 2024
%8 March
%I Association for Computational Linguistics
%C St. Julian’s, Malta
%F hamilton-2024-detecting
%X No two authors write alike. Personal flourishes invoked in written narratives, from lexicon to rhetorical devices, imply a particular author—what literary theorists label the implied or virtual author; distinct from the real author or narrator of a text. Early large language models trained on unfiltered training sets drawn from a variety of discordant sources yielded incoherent personalities, problematic for conversational tasks but proving useful for sampling literature from multiple perspectives. Successes in alignment research in recent years have allowed researchers to impose subjectively consistent personae on language models via instruction tuning and reinforcement learning from human feedback (RLHF), but whether aligned models retain the ability to model an arbitrary virtual author has received little scrutiny. By studying 4,374 stories sampled from three OpenAI language models, we show successive versions of GPT-3 suffer from increasing degrees of “mode collapse” whereby overfitting the model during alignment constrains it from generalizing over authorship: models suffering from mode collapse become unable to assume a multiplicity of perspectives. Our method and results are significant for researchers seeking to employ language models in sociological simulations.
%U https://aclanthology.org/2024.scalellm-1.5
%P 65-72
Markdown (Informal)
[Detecting Mode Collapse in Language Models via Narration](https://aclanthology.org/2024.scalellm-1.5) (Hamilton, SCALE-LLM-WS 2024)
ACL