@inproceedings{tsoneva-etal-2026-scale,
title = "Scale-Invariant Normalization Hides Activation Scaling: Why Patch Site Matters for Demographic Bias Localization in Clinical Text Generation",
author = "Tsoneva, Yoana and
Splitt, Fedor and
Agbaria, Awsam and
Tsiakalou, Alexandra and
Yang, Jing and
Schmitt, Vera",
editor = "Arad, Dana and
Belinkov, Yonatan and
Chen, Hanjie and
Feldhus, Nils and
Mueller, Aaron and
Mohebbi, Hosein and
Sarti, Gabriele and
Tutek, Martin",
booktitle = "Proceedings of the 9th {B}lackbox{NLP} Workshop: Analyzing and Interpreting Neural Networks for {NLP}",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.blackboxnlp-1.66/",
pages = "1155--1166",
ISBN = "979-8-89176-476-7",
abstract = "What activation patching localizes can depend on where the patch is applied. Reproducing Ahsan et al. (2025) and extending it from four to ten models, we find the behavioral results largely reproduce and that medical adaptation improves calibration consistently only for gender with BioMistral, for MedGemma not at all. Two mechanistic findings are artifacts of the patch site. In Gemma-2 and OLMo-2 the patched MLP down-projection passes through a scale-invariant RMSNorm before the residual stream, so activation scaling is ineffective there. Patching after normalization restores it, raising the Gemma-2 flip rate on multiple sclerosis from 0.83, flat under scaling, to 1.00. It also lets a single layer recover the target race nearly as well as the reported sliding window (0.83 vs. 0.84, OLMo-2), so the window compensated for the site, not a distributed representation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tsoneva-etal-2026-scale">
<titleInfo>
<title>Scale-Invariant Normalization Hides Activation Scaling: Why Patch Site Matters for Demographic Bias Localization in Clinical Text Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yoana</namePart>
<namePart type="family">Tsoneva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fedor</namePart>
<namePart type="family">Splitt</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Awsam</namePart>
<namePart type="family">Agbaria</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexandra</namePart>
<namePart type="family">Tsiakalou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jing</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vera</namePart>
<namePart type="family">Schmitt</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 9th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dana</namePart>
<namePart type="family">Arad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yonatan</namePart>
<namePart type="family">Belinkov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanjie</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nils</namePart>
<namePart type="family">Feldhus</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aaron</namePart>
<namePart type="family">Mueller</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hosein</namePart>
<namePart type="family">Mohebbi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gabriele</namePart>
<namePart type="family">Sarti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Martin</namePart>
<namePart type="family">Tutek</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">979-8-89176-476-7</identifier>
</relatedItem>
<abstract>What activation patching localizes can depend on where the patch is applied. Reproducing Ahsan et al. (2025) and extending it from four to ten models, we find the behavioral results largely reproduce and that medical adaptation improves calibration consistently only for gender with BioMistral, for MedGemma not at all. Two mechanistic findings are artifacts of the patch site. In Gemma-2 and OLMo-2 the patched MLP down-projection passes through a scale-invariant RMSNorm before the residual stream, so activation scaling is ineffective there. Patching after normalization restores it, raising the Gemma-2 flip rate on multiple sclerosis from 0.83, flat under scaling, to 1.00. It also lets a single layer recover the target race nearly as well as the reported sliding window (0.83 vs. 0.84, OLMo-2), so the window compensated for the site, not a distributed representation.</abstract>
<identifier type="citekey">tsoneva-etal-2026-scale</identifier>
<location>
<url>https://aclanthology.org/2026.blackboxnlp-1.66/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1155</start>
<end>1166</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Scale-Invariant Normalization Hides Activation Scaling: Why Patch Site Matters for Demographic Bias Localization in Clinical Text Generation
%A Tsoneva, Yoana
%A Splitt, Fedor
%A Agbaria, Awsam
%A Tsiakalou, Alexandra
%A Yang, Jing
%A Schmitt, Vera
%Y Arad, Dana
%Y Belinkov, Yonatan
%Y Chen, Hanjie
%Y Feldhus, Nils
%Y Mueller, Aaron
%Y Mohebbi, Hosein
%Y Sarti, Gabriele
%Y Tutek, Martin
%S Proceedings of the 9th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-476-7
%F tsoneva-etal-2026-scale
%X What activation patching localizes can depend on where the patch is applied. Reproducing Ahsan et al. (2025) and extending it from four to ten models, we find the behavioral results largely reproduce and that medical adaptation improves calibration consistently only for gender with BioMistral, for MedGemma not at all. Two mechanistic findings are artifacts of the patch site. In Gemma-2 and OLMo-2 the patched MLP down-projection passes through a scale-invariant RMSNorm before the residual stream, so activation scaling is ineffective there. Patching after normalization restores it, raising the Gemma-2 flip rate on multiple sclerosis from 0.83, flat under scaling, to 1.00. It also lets a single layer recover the target race nearly as well as the reported sliding window (0.83 vs. 0.84, OLMo-2), so the window compensated for the site, not a distributed representation.
%U https://aclanthology.org/2026.blackboxnlp-1.66/
%P 1155-1166
Markdown (Informal)
[Scale-Invariant Normalization Hides Activation Scaling: Why Patch Site Matters for Demographic Bias Localization in Clinical Text Generation](https://aclanthology.org/2026.blackboxnlp-1.66/) (Tsoneva et al., BlackboxNLP 2026)
ACL