FADE Method Reduces LVLM Hallucinations via FFN Attenuation

FADE is a training-free method that mitigates hallucinations in Large Vision-Language Models by attenuating Feed-Forward Network (FFN) outputs. The approach targets the mechanistic cause of hallucinations where FFN modules at critical layers allow language priors to override visual evidence.