Transformers need glasses! Information over-squashing in language tasks
We study how information propagates in decoder-only Transformers, which are the architectural backbone of most existing frontier large language models (LLMs). We rely on a theoretical signal propagation analysis -- specifically, we analyse the representations of the last token in the final layer of...
Gespeichert in:
Hauptverfasser: | , , , , , , , |
---|---|
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
container_end_page | |
---|---|
container_issue | |
container_start_page | |
container_title | |
container_volume | |
creator | Barbero, Federico Banino, Andrea Kapturowski, Steven Kumaran, Dharshan Araújo, João G. M Vitvitskyi, Alex Pascanu, Razvan Veličković, Petar |
description | We study how information propagates in decoder-only Transformers, which are
the architectural backbone of most existing frontier large language models
(LLMs). We rely on a theoretical signal propagation analysis -- specifically,
we analyse the representations of the last token in the final layer of the
Transformer, as this is the representation used for next-token prediction. Our
analysis reveals a representational collapse phenomenon: we prove that certain
distinct sequences of inputs to the Transformer can yield arbitrarily close
representations in the final token. This effect is exacerbated by the
low-precision floating-point formats frequently used in modern LLMs. As a
result, the model is provably unable to respond to these sequences in different
ways -- leading to errors in, e.g., tasks involving counting or copying.
Further, we show that decoder-only Transformer language models can lose
sensitivity to specific tokens in the input, which relates to the well-known
phenomenon of over-squashing in graph neural networks. We provide empirical
evidence supporting our claims on contemporary LLMs. Our theory also points to
simple solutions towards ameliorating these issues. |
doi_str_mv | 10.48550/arxiv.2406.04267 |
format | Article |
fullrecord | <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2406_04267</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2406_04267</sourcerecordid><originalsourceid>FETCH-arxiv_primary_2406_042673</originalsourceid><addsrcrecordid>eNqFzbsOgkAQheFtLIz6AFaODwCuCGhvvPX0ZBKGdSMMOgNE395A7K1O8ucknzHLrQ3jQ5LYDcrb92EU2zS0cZTup-aSCbKWjdQkCkxUgKtQlXQNNx46tr5haHqSQF8d6t2zA89QIbsOHUGL-tC5mZRYKS1-OzOr8yk7XoORzJ_ia5RPPtD5SO_-P77-Xznw</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Transformers need glasses! Information over-squashing in language tasks</title><source>arXiv.org</source><creator>Barbero, Federico ; Banino, Andrea ; Kapturowski, Steven ; Kumaran, Dharshan ; Araújo, João G. M ; Vitvitskyi, Alex ; Pascanu, Razvan ; Veličković, Petar</creator><creatorcontrib>Barbero, Federico ; Banino, Andrea ; Kapturowski, Steven ; Kumaran, Dharshan ; Araújo, João G. M ; Vitvitskyi, Alex ; Pascanu, Razvan ; Veličković, Petar</creatorcontrib><description>We study how information propagates in decoder-only Transformers, which are
the architectural backbone of most existing frontier large language models
(LLMs). We rely on a theoretical signal propagation analysis -- specifically,
we analyse the representations of the last token in the final layer of the
Transformer, as this is the representation used for next-token prediction. Our
analysis reveals a representational collapse phenomenon: we prove that certain
distinct sequences of inputs to the Transformer can yield arbitrarily close
representations in the final token. This effect is exacerbated by the
low-precision floating-point formats frequently used in modern LLMs. As a
result, the model is provably unable to respond to these sequences in different
ways -- leading to errors in, e.g., tasks involving counting or copying.
Further, we show that decoder-only Transformer language models can lose
sensitivity to specific tokens in the input, which relates to the well-known
phenomenon of over-squashing in graph neural networks. We provide empirical
evidence supporting our claims on contemporary LLMs. Our theory also points to
simple solutions towards ameliorating these issues.</description><identifier>DOI: 10.48550/arxiv.2406.04267</identifier><language>eng</language><subject>Computer Science - Computation and Language ; Computer Science - Learning</subject><creationdate>2024-06</creationdate><rights>http://creativecommons.org/licenses/by/4.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2406.04267$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2406.04267$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Barbero, Federico</creatorcontrib><creatorcontrib>Banino, Andrea</creatorcontrib><creatorcontrib>Kapturowski, Steven</creatorcontrib><creatorcontrib>Kumaran, Dharshan</creatorcontrib><creatorcontrib>Araújo, João G. M</creatorcontrib><creatorcontrib>Vitvitskyi, Alex</creatorcontrib><creatorcontrib>Pascanu, Razvan</creatorcontrib><creatorcontrib>Veličković, Petar</creatorcontrib><title>Transformers need glasses! Information over-squashing in language tasks</title><description>We study how information propagates in decoder-only Transformers, which are
the architectural backbone of most existing frontier large language models
(LLMs). We rely on a theoretical signal propagation analysis -- specifically,
we analyse the representations of the last token in the final layer of the
Transformer, as this is the representation used for next-token prediction. Our
analysis reveals a representational collapse phenomenon: we prove that certain
distinct sequences of inputs to the Transformer can yield arbitrarily close
representations in the final token. This effect is exacerbated by the
low-precision floating-point formats frequently used in modern LLMs. As a
result, the model is provably unable to respond to these sequences in different
ways -- leading to errors in, e.g., tasks involving counting or copying.
Further, we show that decoder-only Transformer language models can lose
sensitivity to specific tokens in the input, which relates to the well-known
phenomenon of over-squashing in graph neural networks. We provide empirical
evidence supporting our claims on contemporary LLMs. Our theory also points to
simple solutions towards ameliorating these issues.</description><subject>Computer Science - Computation and Language</subject><subject>Computer Science - Learning</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNqFzbsOgkAQheFtLIz6AFaODwCuCGhvvPX0ZBKGdSMMOgNE395A7K1O8ucknzHLrQ3jQ5LYDcrb92EU2zS0cZTup-aSCbKWjdQkCkxUgKtQlXQNNx46tr5haHqSQF8d6t2zA89QIbsOHUGL-tC5mZRYKS1-OzOr8yk7XoORzJ_ia5RPPtD5SO_-P77-Xznw</recordid><startdate>20240606</startdate><enddate>20240606</enddate><creator>Barbero, Federico</creator><creator>Banino, Andrea</creator><creator>Kapturowski, Steven</creator><creator>Kumaran, Dharshan</creator><creator>Araújo, João G. M</creator><creator>Vitvitskyi, Alex</creator><creator>Pascanu, Razvan</creator><creator>Veličković, Petar</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20240606</creationdate><title>Transformers need glasses! Information over-squashing in language tasks</title><author>Barbero, Federico ; Banino, Andrea ; Kapturowski, Steven ; Kumaran, Dharshan ; Araújo, João G. M ; Vitvitskyi, Alex ; Pascanu, Razvan ; Veličković, Petar</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-arxiv_primary_2406_042673</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Computer Science - Computation and Language</topic><topic>Computer Science - Learning</topic><toplevel>online_resources</toplevel><creatorcontrib>Barbero, Federico</creatorcontrib><creatorcontrib>Banino, Andrea</creatorcontrib><creatorcontrib>Kapturowski, Steven</creatorcontrib><creatorcontrib>Kumaran, Dharshan</creatorcontrib><creatorcontrib>Araújo, João G. M</creatorcontrib><creatorcontrib>Vitvitskyi, Alex</creatorcontrib><creatorcontrib>Pascanu, Razvan</creatorcontrib><creatorcontrib>Veličković, Petar</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Barbero, Federico</au><au>Banino, Andrea</au><au>Kapturowski, Steven</au><au>Kumaran, Dharshan</au><au>Araújo, João G. M</au><au>Vitvitskyi, Alex</au><au>Pascanu, Razvan</au><au>Veličković, Petar</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Transformers need glasses! Information over-squashing in language tasks</atitle><date>2024-06-06</date><risdate>2024</risdate><abstract>We study how information propagates in decoder-only Transformers, which are
the architectural backbone of most existing frontier large language models
(LLMs). We rely on a theoretical signal propagation analysis -- specifically,
we analyse the representations of the last token in the final layer of the
Transformer, as this is the representation used for next-token prediction. Our
analysis reveals a representational collapse phenomenon: we prove that certain
distinct sequences of inputs to the Transformer can yield arbitrarily close
representations in the final token. This effect is exacerbated by the
low-precision floating-point formats frequently used in modern LLMs. As a
result, the model is provably unable to respond to these sequences in different
ways -- leading to errors in, e.g., tasks involving counting or copying.
Further, we show that decoder-only Transformer language models can lose
sensitivity to specific tokens in the input, which relates to the well-known
phenomenon of over-squashing in graph neural networks. We provide empirical
evidence supporting our claims on contemporary LLMs. Our theory also points to
simple solutions towards ameliorating these issues.</abstract><doi>10.48550/arxiv.2406.04267</doi><oa>free_for_read</oa></addata></record> |
fulltext | fulltext_linktorsrc |
identifier | DOI: 10.48550/arxiv.2406.04267 |
ispartof | |
issn | |
language | eng |
recordid | cdi_arxiv_primary_2406_04267 |
source | arXiv.org |
subjects | Computer Science - Computation and Language Computer Science - Learning |
title | Transformers need glasses! Information over-squashing in language tasks |
url | https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-08T05%3A00%3A25IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Transformers%20need%20glasses!%20Information%20over-squashing%20in%20language%20tasks&rft.au=Barbero,%20Federico&rft.date=2024-06-06&rft_id=info:doi/10.48550/arxiv.2406.04267&rft_dat=%3Carxiv_GOX%3E2406_04267%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true |