IC-457All four tested VLM decoders are heavily text-centric when generating answers, with text modality contributing 85-97% of the prediction signal

Letitia Parcalabescu, Anette Frank

SourceDo Vision & Language Decoders use Images and Text equally? How Self-consistent are their Explanations?

Using an extension of MM-SHAP to autoregressive decoders, the paper measures the multimodal degree of four 7B VLMs. On VQA, t-shap (textual degree) is 87% for bakllava, 97% for llava-next-mistral, 89% for llava-next-vicuna, and 85% for mPLUG-Owl3. On GQA balanced, the values are 90%, 96%, and 90% for the first three models. This means the image modality contributes only 3-15% of the prediction signal, showing that these models predominantly rely on the text input (question, prompt) rather than the image when producing answers.

Evidence
correlational
Key metric
t-shap answer on VQA: bakllava 87%, lv-mistral 97%, lv-vicuna 89%, mplug-owl3 85%; on GQA balanced: bakllava 90%, lv-mistral 96%, lv-vicuna 90%
Caveat
The paper notes this is measured on 100 random samples per dataset due to computational demands. The increase from VQA to GQA balanced is attributed to known stronger linguistic biases in GQA.
Model
BakLLaVA, LLaVA-NeXT / LLaVA 1.6 LLaVA-Next-Mistral, LLaVA-Next-Vicuna, mPLUG-Owl3
Concepts
Shortcut
Datasets
VQA [eval], GQA [eval], GQA Balanced [eval], MS COCO / COCO / COCO 2014 / COCO 2017 / COCO 20k / COCO-it / COCO-wl [eval]
Methods
MM-SHAP [primary]
Related work
MM-SHAP [builds-on]
Related findings
IC-456, IC-458
Extraction
automatic-extraction