@inproceedings{yoo-etal-2026-beyond,
title = "Beyond Diagnosis: Evaluating Recommendation Stability under Counterfactual Skin-Tone Perturbations",
author = "Yoo, Hyunwoo and
Kim, Kyusik and
Rosen, Gail",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.230/",
pages = "4549--4564",
ISBN = "979-8-89176-444-6",
abstract = "Multimodal clinical models have shown strong performance in dermatology recommendation and triage tasks, yet their robustness under demographic perturbations remains poorly understood. We introduce a clinician-validated counterfactual dermatology benchmark for evaluating recommendation stability under skin-tone perturbations in multimodal vision-language models. Our benchmark contains 240 physician-validated images from 60 clinical cases, each paired with four diagnosis- and severity-matched skin-tone variants constructed through iterative LLM filtering and physician review. Across diagnosis, management, urgency, and referral tasks, we observe substantial model-dependent instability even when clinicians verified preservation of diagnosis and severity across perturbation variants. Recommendation changes occur in up to 51.7{\%} of paired comparisons, accompanied by shifts in uncertainty framing, pigmentation-sensitive terminology, and escalation-oriented language. Our findings suggest that demographic perturbations can influence not only diagnostic predictions, but also downstream recommendation behavior in multimodal clinical models. We release a controlled evaluation framework for studying recommendation robustness in medical AI systems."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yoo-etal-2026-beyond">
<titleInfo>
<title>Beyond Diagnosis: Evaluating Recommendation Stability under Counterfactual Skin-Tone Perturbations</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hyunwoo</namePart>
<namePart type="family">Yoo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kyusik</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gail</namePart>
<namePart type="family">Rosen</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>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</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">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Multimodal clinical models have shown strong performance in dermatology recommendation and triage tasks, yet their robustness under demographic perturbations remains poorly understood. We introduce a clinician-validated counterfactual dermatology benchmark for evaluating recommendation stability under skin-tone perturbations in multimodal vision-language models. Our benchmark contains 240 physician-validated images from 60 clinical cases, each paired with four diagnosis- and severity-matched skin-tone variants constructed through iterative LLM filtering and physician review. Across diagnosis, management, urgency, and referral tasks, we observe substantial model-dependent instability even when clinicians verified preservation of diagnosis and severity across perturbation variants. Recommendation changes occur in up to 51.7% of paired comparisons, accompanied by shifts in uncertainty framing, pigmentation-sensitive terminology, and escalation-oriented language. Our findings suggest that demographic perturbations can influence not only diagnostic predictions, but also downstream recommendation behavior in multimodal clinical models. We release a controlled evaluation framework for studying recommendation robustness in medical AI systems.</abstract>
<identifier type="citekey">yoo-etal-2026-beyond</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.230/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>4549</start>
<end>4564</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond Diagnosis: Evaluating Recommendation Stability under Counterfactual Skin-Tone Perturbations
%A Yoo, Hyunwoo
%A Kim, Kyusik
%A Rosen, Gail
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F yoo-etal-2026-beyond
%X Multimodal clinical models have shown strong performance in dermatology recommendation and triage tasks, yet their robustness under demographic perturbations remains poorly understood. We introduce a clinician-validated counterfactual dermatology benchmark for evaluating recommendation stability under skin-tone perturbations in multimodal vision-language models. Our benchmark contains 240 physician-validated images from 60 clinical cases, each paired with four diagnosis- and severity-matched skin-tone variants constructed through iterative LLM filtering and physician review. Across diagnosis, management, urgency, and referral tasks, we observe substantial model-dependent instability even when clinicians verified preservation of diagnosis and severity across perturbation variants. Recommendation changes occur in up to 51.7% of paired comparisons, accompanied by shifts in uncertainty framing, pigmentation-sensitive terminology, and escalation-oriented language. Our findings suggest that demographic perturbations can influence not only diagnostic predictions, but also downstream recommendation behavior in multimodal clinical models. We release a controlled evaluation framework for studying recommendation robustness in medical AI systems.
%U https://aclanthology.org/2026.findings-aacl.230/
%P 4549-4564
Markdown (Informal)
[Beyond Diagnosis: Evaluating Recommendation Stability under Counterfactual Skin-Tone Perturbations](https://aclanthology.org/2026.findings-aacl.230/) (Yoo et al., Findings 2026)
ACL