IC-116In LLaVA-7B, multi-head attention modules drive hallucination more than MLP modules, and targeted intervention on specific hallucination heads reduces the hallucination rate by up to 1.7x
Using zero-ablation (causal mediation analysis) on 1,500 COCO training images, the paper shows that removing MHA modules has approximately 3.5x the influence on hallucination word generation compared to removing MLP modules. Further, a contrastive influence score localizes this effect to specific attention heads. Two intervention strategies—adaptive deactivation of text attention during decoding (AD-HH) and targeted fine-tuning of hallucination heads (TF-HH)—reduce CHAIRs from 51.8 to 29.6 and 35.0 respectively on COCO validation, while maintaining or improving generation quality.
Evidence
interventional
Key metric
MHA influence ~3.5x MLP (Figure 1); CHAIRs 51.8 (greedy) to 29.6 (AD-HH) and 35.0 (TF-HH) on COCO for LLaVA-7B; CHAIRi 13.3 to 8.0 and 8.7; up to 1.7x reduction
Caveat
The hallucination heads were identified on the COCO training set; evaluation is on COCO validation and NoCaps out-of-domain. The method requires explicit attention weight computation, making it incompatible with FlashAttention.