analiza aiqbench
This research reveals a critical failure in current vision-language model alignment strategies, where minimal prompting pressure causes models to default to male representations even when internal encodings retain female associations for ambiguous inputs. The study introduces LALS, a novel zero-shot metric that exposes an asymmetric filtering mechanism within the network layers that systematically suppresses female signals before text generation occurs. These findings highlight significant gaps in how models handle demographic biases under real-world conditions where visual cues are not explicitly clear.
Vision-Language Models Suppress Female Representations Under Ambiguous Input
arXiv:2605.31556v2 Announce Type: replace-cross Abstract: 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 often become 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. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.
Abstract: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 often become 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. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC)
arXiv:2605.31556 [cs.CV]
(or arXiv:2605.31556v2 [cs.CV] for this version)
Text integral sub licența deschisă a publicației · 198 cuvinte
Subiecte conexe
When Visual Evidence is Ambiguous: Pareidolia as a Diagnostic Probe for Vision Models
arXiv · 22 iul. 2026
Compact Robot Policies Need Fine-Grained Visual Representations
arXiv · 7 oct. 2026
Unpacking AI safety for enterprises
Cohere · 4 iun. 2026
Amazon is investing in the Lean Focused Research Organization
Amazon Science · 26 iul. 2026