IC-117Hallucination heads in LLaVA-7B and MiniGPT-4 are concentrated in the middle and deeper layers of the transformer

Tianyun Yang, Ziniu Li, Juan Cao, Chang Xu

SourceUnderstanding and Mitigating Hallucination in Large Vision-Language Models via Modular Attribution and Intervention

After identifying hallucination heads via contrastive influence scoring (top-k selection from 1,024 heads across 32 layers in LLaVA-7B), the paper observes that both the most prominent hallucination heads and non-hallucination heads are predominantly located in the middle and deeper layers. The same pattern is confirmed for MiniGPT-4, where hallucination heads are distributed in the latter half of the model. This aligns with prior findings that deeper layers capture more abstract, task-specific representations.

Evidence
correlational
Key metric
Top 20 hallucination heads (blue boxes) and top non-hallucination heads (red boxes) predominantly in layers 16-32 of 32 (Figure 2); same pattern in MiniGPT-4 (Figure 10)
Caveat
The depth distribution is observed for the top-k heads selected by contrastive influence; the paper does not provide a formal statistical test of the depth distribution.
Model
LLaVA-1.5 / LLaVA-v1.5 LLaVA-v1.5-7B, MiniGPT-4
Concepts
Depth-dependent structure
Datasets
MS COCO / COCO / COCO 2014 / COCO 2017 / COCO 20k / COCO-it / COCO-wl [eval]
Related findings
IC-116, IC-118, IC-119
Extraction
automatic-extraction