@inproceedings{kim-shin-2026-levee,
title = "{LEVEE}: Layer-wise Evaluation of Vector Encoding Effects: Bias Localization and Functional Separation across 11 {LLM}s",
author = "Kim, Dongmin and
Shin, Youhyun",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1394/",
pages = "31845--31874",
ISBN = "979-8-89176-459-0",
abstract = "Large language models encode social biases that instruction fine-tuning aims to suppress. Instruction fine-tuning applies its weight changes through the alignment task vector {\ensuremath{\Delta}}W, yet prior work treats it as a monolithic object, leaving its internal layer structure opaque: which layers carry bias suppression, which carry general capability, and whether these are separable, remains unknown. We map this structure by injecting each layer{'}s {\ensuremath{\Delta}}Wᵢ independently into the base model across 11 LLMs from Qwen2.5, Gemma-2/3, and Llama-3 families (0.5B{--}14B). Testing via per-layer sufficiency sweeps, block removal, and a BBQ{--}MMLU dissociation, we find three results. First, BBQ-measured bias suppression localizes to the mid-third of transformer layers in 9 of 11 models, with mid-third removal raising bias across all 11 while first- and last-third removals produce smaller L-BBQ increases in all nine structured-signal models. Second, mid- and first-third removals dissociate BBQ and MMLU: mid-third removal is BBQ-critical across all tested models while first-third removal can sharply reduce MMLU in Gemma architectures, supporting functional separation of bias suppression and capability in weight space. Third, mid-third removal raises bias above the unaligned base in 3 of 11 cases: partial {\ensuremath{\Delta}}W transfer without the mid-third can make a model more biased than the unaligned base model. These findings provide an empirical roadmap for targeted debiasing and safe task vector composition."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kim-shin-2026-levee">
<titleInfo>
<title>LEVEE: Layer-wise Evaluation of Vector Encoding Effects: Bias Localization and Functional Separation across 11 LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dongmin</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Youhyun</namePart>
<namePart type="family">Shin</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Large language models encode social biases that instruction fine-tuning aims to suppress. Instruction fine-tuning applies its weight changes through the alignment task vector \ensuremathΔW, yet prior work treats it as a monolithic object, leaving its internal layer structure opaque: which layers carry bias suppression, which carry general capability, and whether these are separable, remains unknown. We map this structure by injecting each layer’s \ensuremathΔWᵢ independently into the base model across 11 LLMs from Qwen2.5, Gemma-2/3, and Llama-3 families (0.5B–14B). Testing via per-layer sufficiency sweeps, block removal, and a BBQ–MMLU dissociation, we find three results. First, BBQ-measured bias suppression localizes to the mid-third of transformer layers in 9 of 11 models, with mid-third removal raising bias across all 11 while first- and last-third removals produce smaller L-BBQ increases in all nine structured-signal models. Second, mid- and first-third removals dissociate BBQ and MMLU: mid-third removal is BBQ-critical across all tested models while first-third removal can sharply reduce MMLU in Gemma architectures, supporting functional separation of bias suppression and capability in weight space. Third, mid-third removal raises bias above the unaligned base in 3 of 11 cases: partial \ensuremathΔW transfer without the mid-third can make a model more biased than the unaligned base model. These findings provide an empirical roadmap for targeted debiasing and safe task vector composition.</abstract>
<identifier type="citekey">kim-shin-2026-levee</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1394/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>31845</start>
<end>31874</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LEVEE: Layer-wise Evaluation of Vector Encoding Effects: Bias Localization and Functional Separation across 11 LLMs
%A Kim, Dongmin
%A Shin, Youhyun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F kim-shin-2026-levee
%X Large language models encode social biases that instruction fine-tuning aims to suppress. Instruction fine-tuning applies its weight changes through the alignment task vector \ensuremathΔW, yet prior work treats it as a monolithic object, leaving its internal layer structure opaque: which layers carry bias suppression, which carry general capability, and whether these are separable, remains unknown. We map this structure by injecting each layer’s \ensuremathΔWᵢ independently into the base model across 11 LLMs from Qwen2.5, Gemma-2/3, and Llama-3 families (0.5B–14B). Testing via per-layer sufficiency sweeps, block removal, and a BBQ–MMLU dissociation, we find three results. First, BBQ-measured bias suppression localizes to the mid-third of transformer layers in 9 of 11 models, with mid-third removal raising bias across all 11 while first- and last-third removals produce smaller L-BBQ increases in all nine structured-signal models. Second, mid- and first-third removals dissociate BBQ and MMLU: mid-third removal is BBQ-critical across all tested models while first-third removal can sharply reduce MMLU in Gemma architectures, supporting functional separation of bias suppression and capability in weight space. Third, mid-third removal raises bias above the unaligned base in 3 of 11 cases: partial \ensuremathΔW transfer without the mid-third can make a model more biased than the unaligned base model. These findings provide an empirical roadmap for targeted debiasing and safe task vector composition.
%U https://aclanthology.org/2026.emnlp-main.1394/
%P 31845-31874
Markdown (Informal)
[LEVEE: Layer-wise Evaluation of Vector Encoding Effects: Bias Localization and Functional Separation across 11 LLMs](https://aclanthology.org/2026.emnlp-main.1394/) (Kim & Shin, EMNLP 2026)
ACL