@inproceedings{he-etal-2026-disentangling,
title = "Disentangling Representation Evolution in Transformers through Directional Decomposition",
author = "He, Shwai and
Zhang, Haichao and
Yan, Shen",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1397/",
pages = "29759--29772",
ISBN = "979-8-89176-462-0",
abstract = "Transformer layers evolve representations through learned additive transformations that combine direction-preserving and direction-changing effects. We study this geometry by decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial learned parallel components, showing that direction-preserving behavior is not confined to the residual identity path. To test whether these components are redundant or behaviorally important, we apply the decomposition to residual-space updates and to value aggregation inside attention. Inference-time edits reveal a space-dependent asymmetry: value-space parallel edits preserve behavior more effectively than residual-space or perpendicular counterparts, especially when the direct self message is fixed and only the non-self aggregate is scaled. The decomposition also provides a compact, component-resolved view of compression error. During from-scratch pretraining, using parallel suppression as part of the forward computation lowers validation-loss trajectories and improves downstream averages in our evaluated runs, with the value-space variant strongest. Overall, this geometry connects update direction to editing robustness, compression diagnostics, and training-time intervention. Code is available in the https://github.com/Shwai-He/Transformer-Geometry."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="he-etal-2026-disentangling">
<titleInfo>
<title>Disentangling Representation Evolution in Transformers through Directional Decomposition</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shwai</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haichao</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shen</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Transformer layers evolve representations through learned additive transformations that combine direction-preserving and direction-changing effects. We study this geometry by decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial learned parallel components, showing that direction-preserving behavior is not confined to the residual identity path. To test whether these components are redundant or behaviorally important, we apply the decomposition to residual-space updates and to value aggregation inside attention. Inference-time edits reveal a space-dependent asymmetry: value-space parallel edits preserve behavior more effectively than residual-space or perpendicular counterparts, especially when the direct self message is fixed and only the non-self aggregate is scaled. The decomposition also provides a compact, component-resolved view of compression error. During from-scratch pretraining, using parallel suppression as part of the forward computation lowers validation-loss trajectories and improves downstream averages in our evaluated runs, with the value-space variant strongest. Overall, this geometry connects update direction to editing robustness, compression diagnostics, and training-time intervention. Code is available in the https://github.com/Shwai-He/Transformer-Geometry.</abstract>
<identifier type="citekey">he-etal-2026-disentangling</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1397/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>29759</start>
<end>29772</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Disentangling Representation Evolution in Transformers through Directional Decomposition
%A He, Shwai
%A Zhang, Haichao
%A Yan, Shen
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F he-etal-2026-disentangling
%X Transformer layers evolve representations through learned additive transformations that combine direction-preserving and direction-changing effects. We study this geometry by decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial learned parallel components, showing that direction-preserving behavior is not confined to the residual identity path. To test whether these components are redundant or behaviorally important, we apply the decomposition to residual-space updates and to value aggregation inside attention. Inference-time edits reveal a space-dependent asymmetry: value-space parallel edits preserve behavior more effectively than residual-space or perpendicular counterparts, especially when the direct self message is fixed and only the non-self aggregate is scaled. The decomposition also provides a compact, component-resolved view of compression error. During from-scratch pretraining, using parallel suppression as part of the forward computation lowers validation-loss trajectories and improves downstream averages in our evaluated runs, with the value-space variant strongest. Overall, this geometry connects update direction to editing robustness, compression diagnostics, and training-time intervention. Code is available in the https://github.com/Shwai-He/Transformer-Geometry.
%U https://aclanthology.org/2026.findings-emnlp.1397/
%P 29759-29772
Markdown (Informal)
[Disentangling Representation Evolution in Transformers through Directional Decomposition](https://aclanthology.org/2026.findings-emnlp.1397/) (He et al., Findings 2026)
ACL