Vision-Language Models Suppress Female Representations Under Ambiguous Input
A new research paper reveals that vision-language models (VLMs) exhibit a hidden bias against female representations, even when aligned to avoid demographic stereotypes. When presented with ambiguous visual inputs, these models tend to default to male associations for occupations, regardless of their internal encoding. The study introduces a metric called LALS to measure these internal concept associations, finding that female signals are suppressed before output generation, unlike male signals which amplify throughout the process. AI
IMPACT Reveals a subtle, internal bias in VLMs that may affect downstream applications and highlights the need for more nuanced alignment techniques.