StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs

arXiv·medium signal

Kolli, Cavelius, and Nikeghbal find that a small set of human visual cues accounts for the majority of social biases exhibited by multimodal LLMs deployed in consequential settings. The implication is that targeted interventions on these few cues could mitigate bias more efficiently than broad debiasing. Relevant for teams deploying MLLMs where fairness matters.