@inproceedings{havstein-sogaard-2026-alignment,
title = "Does Alignment Debias or Just Suppress? Evaluating Stereotypical Bias Across Base-Instruct Pairs",
author = "Havstein, Sigurd Kornelius and
S{\o}gaard, Anders",
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.1199/",
pages = "27446--27460",
ISBN = "979-8-89176-459-0",
abstract = "Alignment has been widely reported to reduce stereotypical bias in large language models, but it is unclear whether these reductions reflect debiasing or suppression at the output level. Across 27 base{--}instruct pairs from four open-weight families and four bias benchmarks, three patterns provide substantial evidence that the measured reduction is largely suppression. (i) The chat template drives about half the alignment effect on CrowS-Pairs and essentially all on StereoSet, where 18 of 27 improve without it (p = 0.061) versus 26 of 27 with it (p {\ensuremath{<}} 0.001). Per-pair effects remain below d = 0.2 in both conditions. IAT lies at the weight-level end of the spectrum, with template contribution indistinguishable from zero. (ii) Decomposing BBQ{'}s ambiguous-context score shows instruct models defer rather than debias (28{\%} {\textrightarrow} 53{\%}) and are more biased when they answer (conditional bias 0.53 {\textrightarrow} 0.59). Suppressed bias is recoverable: the most aggressive adversarial framing restores CrowS-Pairs bias to near base levels, while on BBQ it partly reverses deferral without reactivating conditional bias. (iii) Across eight probed pairs, base and instruct gender directions remain geometrically aligned (median cosine 0.86 vs. 0.006 random-pair baseline). Together, these findings suggest benchmark improvements after alignment may overstate genuine bias reduction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="havstein-sogaard-2026-alignment">
<titleInfo>
<title>Does Alignment Debias or Just Suppress? Evaluating Stereotypical Bias Across Base-Instruct Pairs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sigurd</namePart>
<namePart type="given">Kornelius</namePart>
<namePart type="family">Havstein</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anders</namePart>
<namePart type="family">Søgaard</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>Alignment has been widely reported to reduce stereotypical bias in large language models, but it is unclear whether these reductions reflect debiasing or suppression at the output level. Across 27 base–instruct pairs from four open-weight families and four bias benchmarks, three patterns provide substantial evidence that the measured reduction is largely suppression. (i) The chat template drives about half the alignment effect on CrowS-Pairs and essentially all on StereoSet, where 18 of 27 improve without it (p = 0.061) versus 26 of 27 with it (p \ensuremath< 0.001). Per-pair effects remain below d = 0.2 in both conditions. IAT lies at the weight-level end of the spectrum, with template contribution indistinguishable from zero. (ii) Decomposing BBQ’s ambiguous-context score shows instruct models defer rather than debias (28% → 53%) and are more biased when they answer (conditional bias 0.53 → 0.59). Suppressed bias is recoverable: the most aggressive adversarial framing restores CrowS-Pairs bias to near base levels, while on BBQ it partly reverses deferral without reactivating conditional bias. (iii) Across eight probed pairs, base and instruct gender directions remain geometrically aligned (median cosine 0.86 vs. 0.006 random-pair baseline). Together, these findings suggest benchmark improvements after alignment may overstate genuine bias reduction.</abstract>
<identifier type="citekey">havstein-sogaard-2026-alignment</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1199/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>27446</start>
<end>27460</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Does Alignment Debias or Just Suppress? Evaluating Stereotypical Bias Across Base-Instruct Pairs
%A Havstein, Sigurd Kornelius
%A Søgaard, Anders
%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 havstein-sogaard-2026-alignment
%X Alignment has been widely reported to reduce stereotypical bias in large language models, but it is unclear whether these reductions reflect debiasing or suppression at the output level. Across 27 base–instruct pairs from four open-weight families and four bias benchmarks, three patterns provide substantial evidence that the measured reduction is largely suppression. (i) The chat template drives about half the alignment effect on CrowS-Pairs and essentially all on StereoSet, where 18 of 27 improve without it (p = 0.061) versus 26 of 27 with it (p \ensuremath< 0.001). Per-pair effects remain below d = 0.2 in both conditions. IAT lies at the weight-level end of the spectrum, with template contribution indistinguishable from zero. (ii) Decomposing BBQ’s ambiguous-context score shows instruct models defer rather than debias (28% → 53%) and are more biased when they answer (conditional bias 0.53 → 0.59). Suppressed bias is recoverable: the most aggressive adversarial framing restores CrowS-Pairs bias to near base levels, while on BBQ it partly reverses deferral without reactivating conditional bias. (iii) Across eight probed pairs, base and instruct gender directions remain geometrically aligned (median cosine 0.86 vs. 0.006 random-pair baseline). Together, these findings suggest benchmark improvements after alignment may overstate genuine bias reduction.
%U https://aclanthology.org/2026.emnlp-main.1199/
%P 27446-27460
Markdown (Informal)
[Does Alignment Debias or Just Suppress? Evaluating Stereotypical Bias Across Base-Instruct Pairs](https://aclanthology.org/2026.emnlp-main.1199/) (Havstein & Søgaard, EMNLP 2026)
ACL