StylisticBias: 15 Visual Attributes Account for 80% of Social Bias in Multimodal LLMs

Research official 1 src. ~1 min

A controlled benchmark of ~25,000 photorealistic images — ~50 per-attribute variations per base face with identity held constant — shows that age and body type dominate identity-level bias in MLLMs, while fashion style drives the largest attribute-level shifts. Across six MLLMs and 25 social judgment scenarios, ~15 attributes account for ~80% of total bias variation. Accepted to ICML 2026 workshops.

Why it matters

Provides a Pareto account of MLLM social bias: practitioners can focus on a small high-leverage set of visual attributes rather than auditing all possible variables. The methodology of isolating attributes with identity constant is cleaner than prior holistic evaluations.

Importance: 2/5

Solid ICML workshop paper with actionable findings for MLLM bias practitioners.

Sources