Vision-Language Models Suppress Female Representations Under Ambiguous Input

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Marin-Llobet, Arnau, Henniger, Simon, Banaji, Mahzarin R.
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866910272568950784
author Marin-Llobet, Arnau
Henniger, Simon
Banaji, Mahzarin R.
author_facet Marin-Llobet, Arnau
Henniger, Simon
Banaji, Mahzarin R.
contents Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind) cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs are systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter -- male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation -- and a color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.
format Preprint
id arxiv_https___arxiv_org_abs_2605_31556
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Vision-Language Models Suppress Female Representations Under Ambiguous Input
Marin-Llobet, Arnau
Henniger, Simon
Banaji, Mahzarin R.
Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Computers and Society
Human-Computer Interaction
Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind) cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs are systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter -- male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation -- and a color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.
title Vision-Language Models Suppress Female Representations Under Ambiguous Input
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Computers and Society
Human-Computer Interaction
url https://arxiv.org/abs/2605.31556