Irrelevant Text Causes Affine Margin Shifts in MLLMs
August 21, 2026
Multimodal LLMs exhibit systematic bias from task-irrelevant context rather than stochastic noise. The study finds that auxiliary text shifts decision margins following a consistent affine transformation, allowing for predictable estimation of context-induced distortion.
HOW THIS AFFECTS YOU
●
researcherYou can use these geometric regularity findings to better model and mitigate context-induced bias in multimodal systems.