The study investigates whether object-specific information in the visual tokens of the language model component of LLaVA-1.5 is localized or dispersed. Ablation experiments were conducted where tokens corresponding to the object's spatial location were removed and replaced with a mean embedding. This led to a significant drop in the model's ability to identify the object across three tasks (generative description, binary polling, and VQA). The drop was substantially larger than when ablating an equivalent number of random tokens or tokens identified as important via Integrated Gradients, supporting the claim of localization.
Evidence
interventional
Key metric
For generative description: 71.79% decrease (for +1 buffer, 33.4 tokens); For polling: 51.28% decrease; For VQA: 86.67% decrease.
Caveat
The results are for object identification tasks and on the LLaVA family of models. The study only investigates a subset of 100 images for the VQA task, and the filtering process for the dataset may introduce bias.