Tag
This paper identifies imbalanced attention head groups in MLLMs that drive or resist modality-conflict hallucination, and proposes MACI, a causal intervention that suppresses hallucination-driving heads only when conflict is detected, achieving large hallucination reduction across five models.