@inproceedings{xu-etal-2026-progressive,
title = "Progressive Behavioral Drift through Compression Valleys in Large Language Models",
author = "Xu, Zhiyuan and
Abaimov, Stanislav and
Gardiner, Joseph and
Belguith, Sana",
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.1967/",
pages = "41941--41964",
ISBN = "979-8-89176-462-0",
abstract = "We show that attention sinks and compression valleys create a vulnerable region in decoder-only Transformers, where small activation perturbations can be amplified through the autoregressive trajectory. Based on this, we propose Sensitivity-Scaled Steering (SSS), a progressive activation-space attack that anchors perturbations at the beginning-of-sequence token and adaptively reinforces them at sensitive layers and tokens. Instead of forcing an abrupt behavioral change, SSS induces a staged drift, making outputs gradually shift toward the target behavior while remaining fluent and benign-looking in early generations. Across multiple open-weight models and four behavioral axes, SSS achieves high attack success, preserves coherence, and causes negligible degradation to general capabilities. These results show that attention sinks and compression valleys are not merely mechanistic features; rather, they expose exploitable amplification mechanisms that can be treated as hidden-state weaknesses for activation-space attacks in white-box and supply-chain LLM deployments."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-progressive">
<titleInfo>
<title>Progressive Behavioral Drift through Compression Valleys in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhiyuan</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stanislav</namePart>
<namePart type="family">Abaimov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joseph</namePart>
<namePart type="family">Gardiner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sana</namePart>
<namePart type="family">Belguith</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>We show that attention sinks and compression valleys create a vulnerable region in decoder-only Transformers, where small activation perturbations can be amplified through the autoregressive trajectory. Based on this, we propose Sensitivity-Scaled Steering (SSS), a progressive activation-space attack that anchors perturbations at the beginning-of-sequence token and adaptively reinforces them at sensitive layers and tokens. Instead of forcing an abrupt behavioral change, SSS induces a staged drift, making outputs gradually shift toward the target behavior while remaining fluent and benign-looking in early generations. Across multiple open-weight models and four behavioral axes, SSS achieves high attack success, preserves coherence, and causes negligible degradation to general capabilities. These results show that attention sinks and compression valleys are not merely mechanistic features; rather, they expose exploitable amplification mechanisms that can be treated as hidden-state weaknesses for activation-space attacks in white-box and supply-chain LLM deployments.</abstract>
<identifier type="citekey">xu-etal-2026-progressive</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1967/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>41941</start>
<end>41964</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Progressive Behavioral Drift through Compression Valleys in Large Language Models
%A Xu, Zhiyuan
%A Abaimov, Stanislav
%A Gardiner, Joseph
%A Belguith, Sana
%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 xu-etal-2026-progressive
%X We show that attention sinks and compression valleys create a vulnerable region in decoder-only Transformers, where small activation perturbations can be amplified through the autoregressive trajectory. Based on this, we propose Sensitivity-Scaled Steering (SSS), a progressive activation-space attack that anchors perturbations at the beginning-of-sequence token and adaptively reinforces them at sensitive layers and tokens. Instead of forcing an abrupt behavioral change, SSS induces a staged drift, making outputs gradually shift toward the target behavior while remaining fluent and benign-looking in early generations. Across multiple open-weight models and four behavioral axes, SSS achieves high attack success, preserves coherence, and causes negligible degradation to general capabilities. These results show that attention sinks and compression valleys are not merely mechanistic features; rather, they expose exploitable amplification mechanisms that can be treated as hidden-state weaknesses for activation-space attacks in white-box and supply-chain LLM deployments.
%U https://aclanthology.org/2026.findings-emnlp.1967/
%P 41941-41964
Markdown (Informal)
[Progressive Behavioral Drift through Compression Valleys in Large Language Models](https://aclanthology.org/2026.findings-emnlp.1967/) (Xu et al., Findings 2026)
ACL