IC-020LLaVA-1.5 extracts object information directly from visual tokens to the last token in mid-late layers

Clement Neo, Luke Ong, Philip Torr, Mor Geva, David Krueger, Fazl Barez

SourceTowards Interpreting Visual Information Processing in Vision-Language Models

The study investigated how visual information flows to the last token position for object identification. Using attention knockout, they selectively blocked attention connections between different token groups across layers. Blocking attention from the object tokens and their surrounding buffer to the last token position in mid-late layers led to a significant drop in prediction accuracy and probability. In contrast, blocking attention from visual tokens to the last row of visual tokens had minimal effect, suggesting the model does not rely on a summarization step in this last row for the task. These findings were replicated with a sliding window on LLaVA-Phi-3.

Evidence
interventional
Key metric
Relative performance on the correct token drops to 0.88 (accuracy) and 0.82 (probability) when blocking attention from object tokens to the last token position in mid-late layers.
Caveat
The analysis focuses on a simple object identification task and may not generalize to more complex reasoning tasks.
Model
LLaVA-1.5 / LLaVA-v1.5, LLaVA-Phi LLaVA-Phi-3
Concepts
Depth-dependent structure
Datasets
MS COCO / COCO / COCO 2014 / COCO 2017 / COCO 20k / COCO-it / COCO-wl [eval]
Methods
Attention knockout [primary]
Related work
Basu et al. (2024) [compared-to]
Extraction
automatic-extraction