@inproceedings{nooralahzadeh-etal-2026-universal,
title = "Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models",
author = "Nooralahzadeh, Farhad and
Gundersen, Benjamin and
Deperrois, Nicolas and
Matsuo, Hidetoshi and
Nishio, Mizuho and
Frauenfelder, Thomas and
Allam, Ahmed and
Bluethgen, Christian and
Moor, Michael and
Krauthammer, Michael",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.102/",
pages = "2241--2263",
ISBN = "979-8-89176-439-2",
abstract = "Medical vision-language models (VLMs) commit clinical errors when generating chest X-ray reports: they fabricate findings that are not present in the image, miss important ones, or locate them incorrectly. We steer the error composition of generated reports \textit{without} weight updates by decoding-time residual steering on a per-token sparse autoencoder (SAE) basis: Top-$K$ SAEs on late layers, per-error-type causal feature screening against clinical errors, then combined \textit{suppress/boost} intervention at inference time. On the MIMIC-CXR test split, our inference-only method improves the quality of generated reports for three radiology VLMs (RadVLM, LLaVA-Rad, and CheXOne), with relative improvements of +5.4{\%}, +7.2{\%}, and +17.0{\%} in the clinical composite metric, and statistically significant GREEN gains on all backbones. The gain is an explicit, controllable trade-off rather than uniform error reduction: steering shifts models toward \textit{more complete reporting}, substantially fewer missing findings, at the cost of some additional false findings, and we analyse and quantify this redistribution per error type. A cross-model feature alignment shows that the quality-promoting (boost) directions have functionally similar error-type profiles across architectures, whereas hallucination-linkedm (suppress) directions are model-specific. Census-motivated steering should therefore treat suppression per-backbone, rather than sharing a universal suppress list. The same recipe applied zero-shot to IU-Xray (GREEN $+7.7\%${~}rel., on RadVLM) suggests that the identified features are properties of the model rather than of the training corpus. The causal feature sets and an interactive feature dashboard are shared."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nooralahzadeh-etal-2026-universal">
<titleInfo>
<title>Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Farhad</namePart>
<namePart type="family">Nooralahzadeh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Benjamin</namePart>
<namePart type="family">Gundersen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nicolas</namePart>
<namePart type="family">Deperrois</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hidetoshi</namePart>
<namePart type="family">Matsuo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mizuho</namePart>
<namePart type="family">Nishio</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thomas</namePart>
<namePart type="family">Frauenfelder</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ahmed</namePart>
<namePart type="family">Allam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christian</namePart>
<namePart type="family">Bluethgen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Moor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Krauthammer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Medical vision-language models (VLMs) commit clinical errors when generating chest X-ray reports: they fabricate findings that are not present in the image, miss important ones, or locate them incorrectly. We steer the error composition of generated reports without weight updates by decoding-time residual steering on a per-token sparse autoencoder (SAE) basis: Top-K SAEs on late layers, per-error-type causal feature screening against clinical errors, then combined suppress/boost intervention at inference time. On the MIMIC-CXR test split, our inference-only method improves the quality of generated reports for three radiology VLMs (RadVLM, LLaVA-Rad, and CheXOne), with relative improvements of +5.4%, +7.2%, and +17.0% in the clinical composite metric, and statistically significant GREEN gains on all backbones. The gain is an explicit, controllable trade-off rather than uniform error reduction: steering shifts models toward more complete reporting, substantially fewer missing findings, at the cost of some additional false findings, and we analyse and quantify this redistribution per error type. A cross-model feature alignment shows that the quality-promoting (boost) directions have functionally similar error-type profiles across architectures, whereas hallucination-linkedm (suppress) directions are model-specific. Census-motivated steering should therefore treat suppression per-backbone, rather than sharing a universal suppress list. The same recipe applied zero-shot to IU-Xray (GREEN +7.7% rel., on RadVLM) suggests that the identified features are properties of the model rather than of the training corpus. The causal feature sets and an interactive feature dashboard are shared.</abstract>
<identifier type="citekey">nooralahzadeh-etal-2026-universal</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.102/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>2241</start>
<end>2263</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models
%A Nooralahzadeh, Farhad
%A Gundersen, Benjamin
%A Deperrois, Nicolas
%A Matsuo, Hidetoshi
%A Nishio, Mizuho
%A Frauenfelder, Thomas
%A Allam, Ahmed
%A Bluethgen, Christian
%A Moor, Michael
%A Krauthammer, Michael
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F nooralahzadeh-etal-2026-universal
%X Medical vision-language models (VLMs) commit clinical errors when generating chest X-ray reports: they fabricate findings that are not present in the image, miss important ones, or locate them incorrectly. We steer the error composition of generated reports without weight updates by decoding-time residual steering on a per-token sparse autoencoder (SAE) basis: Top-K SAEs on late layers, per-error-type causal feature screening against clinical errors, then combined suppress/boost intervention at inference time. On the MIMIC-CXR test split, our inference-only method improves the quality of generated reports for three radiology VLMs (RadVLM, LLaVA-Rad, and CheXOne), with relative improvements of +5.4%, +7.2%, and +17.0% in the clinical composite metric, and statistically significant GREEN gains on all backbones. The gain is an explicit, controllable trade-off rather than uniform error reduction: steering shifts models toward more complete reporting, substantially fewer missing findings, at the cost of some additional false findings, and we analyse and quantify this redistribution per error type. A cross-model feature alignment shows that the quality-promoting (boost) directions have functionally similar error-type profiles across architectures, whereas hallucination-linkedm (suppress) directions are model-specific. Census-motivated steering should therefore treat suppression per-backbone, rather than sharing a universal suppress list. The same recipe applied zero-shot to IU-Xray (GREEN +7.7% rel., on RadVLM) suggests that the identified features are properties of the model rather than of the training corpus. The causal feature sets and an interactive feature dashboard are shared.
%U https://aclanthology.org/2026.aacl-long.102/
%P 2241-2263
Markdown (Informal)
[Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models](https://aclanthology.org/2026.aacl-long.102/) (Nooralahzadeh et al., AACL-IJCNLP 2026)
ACL
- Farhad Nooralahzadeh, Benjamin Gundersen, Nicolas Deperrois, Hidetoshi Matsuo, Mizuho Nishio, Thomas Frauenfelder, Ahmed Allam, Christian Bluethgen, Michael Moor, and Michael Krauthammer. 2026. Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2241–2263, Hengqin, China. Association for Computational Linguistics.