@inproceedings{bertolini-etal-2025-strengths,
title = "Strengths and Limitations of Word-Based Task Explainability in Vision Language Models: a Case Study on Biological Sex Biases in the Medical Domain",
author = "Bertolini, Lorenzo and
Comte, Valentin and
Ruiz-Serra, Victoria and
Orfei, Lia and
Ceresa, Mario",
editor = "Fale{\'n}ska, Agnieszka and
Basta, Christine and
Costa-juss{\`a}, Marta and
Sta{\'n}czak, Karolina and
Nozza, Debora",
booktitle = "Proceedings of the 6th Workshop on Gender Bias in Natural Language Processing (GeBNLP)",
month = aug,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.gebnlp-1.12/",
doi = "10.18653/v1/2025.gebnlp-1.12",
pages = "111--123",
ISBN = "979-8-89176-277-0",
abstract = "Vision-language models (VLMs) can achieve high accuracy in medical applications but can retain demographic biases from training data. While multiple works have identified the presence of these biases in many VLMs, it remains unclear how strong their impact at the inference level is. In this work, we study how well a task-level explainability method based on linear combinations of words can detect multiple types of biases, with a focus on medical image classification. By manipulating the training datasets with demographic and non-demographic biases, we show how the adopted approach can detect explicitly encoded biases but fails with implicitly encoded ones, particularly biological sex. Our results suggest that such a failure likely stems from misalignment between sex-describing features in image versus text modalities. Our findings highlight limitations in the evaluated explainability method for detecting implicit biases in medical VLMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bertolini-etal-2025-strengths">
<titleInfo>
<title>Strengths and Limitations of Word-Based Task Explainability in Vision Language Models: a Case Study on Biological Sex Biases in the Medical Domain</title>
</titleInfo>
<name type="personal">
<namePart type="given">Lorenzo</namePart>
<namePart type="family">Bertolini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Valentin</namePart>
<namePart type="family">Comte</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Victoria</namePart>
<namePart type="family">Ruiz-Serra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lia</namePart>
<namePart type="family">Orfei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mario</namePart>
<namePart type="family">Ceresa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 6th Workshop on Gender Bias in Natural Language Processing (GeBNLP)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Agnieszka</namePart>
<namePart type="family">Faleńska</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christine</namePart>
<namePart type="family">Basta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marta</namePart>
<namePart type="family">Costa-jussà</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Karolina</namePart>
<namePart type="family">Stańczak</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Debora</namePart>
<namePart type="family">Nozza</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Vienna, Austria</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-277-0</identifier>
</relatedItem>
<abstract>Vision-language models (VLMs) can achieve high accuracy in medical applications but can retain demographic biases from training data. While multiple works have identified the presence of these biases in many VLMs, it remains unclear how strong their impact at the inference level is. In this work, we study how well a task-level explainability method based on linear combinations of words can detect multiple types of biases, with a focus on medical image classification. By manipulating the training datasets with demographic and non-demographic biases, we show how the adopted approach can detect explicitly encoded biases but fails with implicitly encoded ones, particularly biological sex. Our results suggest that such a failure likely stems from misalignment between sex-describing features in image versus text modalities. Our findings highlight limitations in the evaluated explainability method for detecting implicit biases in medical VLMs.</abstract>
<identifier type="citekey">bertolini-etal-2025-strengths</identifier>
<identifier type="doi">10.18653/v1/2025.gebnlp-1.12</identifier>
<location>
<url>https://aclanthology.org/2025.gebnlp-1.12/</url>
</location>
<part>
<date>2025-08</date>
<extent unit="page">
<start>111</start>
<end>123</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Strengths and Limitations of Word-Based Task Explainability in Vision Language Models: a Case Study on Biological Sex Biases in the Medical Domain
%A Bertolini, Lorenzo
%A Comte, Valentin
%A Ruiz-Serra, Victoria
%A Orfei, Lia
%A Ceresa, Mario
%Y Faleńska, Agnieszka
%Y Basta, Christine
%Y Costa-jussà, Marta
%Y Stańczak, Karolina
%Y Nozza, Debora
%S Proceedings of the 6th Workshop on Gender Bias in Natural Language Processing (GeBNLP)
%D 2025
%8 August
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-277-0
%F bertolini-etal-2025-strengths
%X Vision-language models (VLMs) can achieve high accuracy in medical applications but can retain demographic biases from training data. While multiple works have identified the presence of these biases in many VLMs, it remains unclear how strong their impact at the inference level is. In this work, we study how well a task-level explainability method based on linear combinations of words can detect multiple types of biases, with a focus on medical image classification. By manipulating the training datasets with demographic and non-demographic biases, we show how the adopted approach can detect explicitly encoded biases but fails with implicitly encoded ones, particularly biological sex. Our results suggest that such a failure likely stems from misalignment between sex-describing features in image versus text modalities. Our findings highlight limitations in the evaluated explainability method for detecting implicit biases in medical VLMs.
%R 10.18653/v1/2025.gebnlp-1.12
%U https://aclanthology.org/2025.gebnlp-1.12/
%U https://doi.org/10.18653/v1/2025.gebnlp-1.12
%P 111-123
Markdown (Informal)
[Strengths and Limitations of Word-Based Task Explainability in Vision Language Models: a Case Study on Biological Sex Biases in the Medical Domain](https://aclanthology.org/2025.gebnlp-1.12/) (Bertolini et al., GeBNLP 2025)
ACL