@inproceedings{faizan-etal-2026-auditing,
title = "Auditing and Mitigating Paraphrase Fragility in Medical Vision-Language Models",
author = "Faizan, Syed and
Chandramohan, Anuradha and
Krishnamurthy, Parameswari",
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.549/",
pages = "11575--11592",
ISBN = "979-8-89176-462-0",
abstract = "Paraphrases express the same intent in different words, so a robust vision-language model (VLM) should preserve its clinical conclusion when a question is rephrased. Medical VLMs often violate this expectation. We audit this failure on VQA-RAD using nine clinically grounded linguistic operations and 9,723 answer-preserving paraphrases. Across eight checkpoints from five architecture stacks (2.2B{--}27.4B), Negation Reframing is consistently the most fragile category, with clinical flip rates from 30.7{\%} to 95.3{\%}. A locked, category-conditioned LoRA reduces this failure by 59.0{--}75.8{\%} across all eight checkpoints; on MedGemma-27B, paraphrase accuracy rises from 68.4{\%} to 75.5{\%}, although base accuracy falls from 79.6{\%} to 73.9{\%}. A pooled adapter performs nearly as well, so we use category-conditioned adapters primarily as diagnostic instruments. Two clinicians agree on 97.8{\%} of a 500-pair calibration sample ({\ensuremath{\kappa}}=0.901), whereas agreement with an LLM equivalence judge is only {\ensuremath{\kappa}}{\ensuremath{\approx}}0.14. Sparse-autoencoder clamping yields only a 5.8{\%} reduction, and activation patching provides suggestive rather than conclusive localization evidence. Together, these results provide a reproducible robustness audit for radiology VLMs and make its dataset, generator, evaluation, and deployment limits explicit."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="faizan-etal-2026-auditing">
<titleInfo>
<title>Auditing and Mitigating Paraphrase Fragility in Medical Vision-Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Syed</namePart>
<namePart type="family">Faizan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anuradha</namePart>
<namePart type="family">Chandramohan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Parameswari</namePart>
<namePart type="family">Krishnamurthy</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>Paraphrases express the same intent in different words, so a robust vision-language model (VLM) should preserve its clinical conclusion when a question is rephrased. Medical VLMs often violate this expectation. We audit this failure on VQA-RAD using nine clinically grounded linguistic operations and 9,723 answer-preserving paraphrases. Across eight checkpoints from five architecture stacks (2.2B–27.4B), Negation Reframing is consistently the most fragile category, with clinical flip rates from 30.7% to 95.3%. A locked, category-conditioned LoRA reduces this failure by 59.0–75.8% across all eight checkpoints; on MedGemma-27B, paraphrase accuracy rises from 68.4% to 75.5%, although base accuracy falls from 79.6% to 73.9%. A pooled adapter performs nearly as well, so we use category-conditioned adapters primarily as diagnostic instruments. Two clinicians agree on 97.8% of a 500-pair calibration sample (\ensuremathąppa=0.901), whereas agreement with an LLM equivalence judge is only \ensuremathąppa\ensuremath\approx0.14. Sparse-autoencoder clamping yields only a 5.8% reduction, and activation patching provides suggestive rather than conclusive localization evidence. Together, these results provide a reproducible robustness audit for radiology VLMs and make its dataset, generator, evaluation, and deployment limits explicit.</abstract>
<identifier type="citekey">faizan-etal-2026-auditing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.549/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>11575</start>
<end>11592</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Auditing and Mitigating Paraphrase Fragility in Medical Vision-Language Models
%A Faizan, Syed
%A Chandramohan, Anuradha
%A Krishnamurthy, Parameswari
%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 faizan-etal-2026-auditing
%X Paraphrases express the same intent in different words, so a robust vision-language model (VLM) should preserve its clinical conclusion when a question is rephrased. Medical VLMs often violate this expectation. We audit this failure on VQA-RAD using nine clinically grounded linguistic operations and 9,723 answer-preserving paraphrases. Across eight checkpoints from five architecture stacks (2.2B–27.4B), Negation Reframing is consistently the most fragile category, with clinical flip rates from 30.7% to 95.3%. A locked, category-conditioned LoRA reduces this failure by 59.0–75.8% across all eight checkpoints; on MedGemma-27B, paraphrase accuracy rises from 68.4% to 75.5%, although base accuracy falls from 79.6% to 73.9%. A pooled adapter performs nearly as well, so we use category-conditioned adapters primarily as diagnostic instruments. Two clinicians agree on 97.8% of a 500-pair calibration sample (\ensuremathąppa=0.901), whereas agreement with an LLM equivalence judge is only \ensuremathąppa\ensuremath\approx0.14. Sparse-autoencoder clamping yields only a 5.8% reduction, and activation patching provides suggestive rather than conclusive localization evidence. Together, these results provide a reproducible robustness audit for radiology VLMs and make its dataset, generator, evaluation, and deployment limits explicit.
%U https://aclanthology.org/2026.findings-emnlp.549/
%P 11575-11592
Markdown (Informal)
[Auditing and Mitigating Paraphrase Fragility in Medical Vision-Language Models](https://aclanthology.org/2026.findings-emnlp.549/) (Faizan et al., Findings 2026)
ACL