The paper evaluates the visual interpretability of LLaVA by examining the composite attention maps (integrating LLM and ViT attention weights) over the input image for tokens in the generated sentence. Using the original CLIP encoder, LLaVA's attention maps are noisy and fail to sparsely highlight the relevant objects. Quantitatively, on 144 overlapping images between ImageNetette and ImageNet-Segmentation, the original CLIP encoder yields a Gini index of 25.29, pointing game of 5.56, and mask-IoU of 4.13. The paper also shows qualitatively (Figure 7) that attention maps with the original CLIP are diffuse, while AFT-CLIP produces cleaner, more object-focused maps.
Evidence
correlational
Key metric
Gini 25.29, PG 5.56, PG-energy 19.71, pixel acc. 69.62, AP 28.77, mask-IoU 4.13 (original CLIP encoder in LLaVA, 144 images)
Caveat
The authors note that ViT attention maps frequently contain high-norm tokens in low-informative background areas (a well-documented issue), which significantly impacts the proposed metrics and is not fully mitigated by AFT.