ViT-CX: Causal Explanation of Vision Transformers

Existing explanation methods for Vision Transformers are inadequate: Current approaches that rely on attention weights often produce "unsatisfactory saliency maps," meaning they fail to clearly show which parts of an image a model actually uses to make decisions.

ViT-CX uses a fundamentally different approach: Rather than tracking where the model pays attention, this new method analyzes patch embeddings (the model's internal representations of image sections) and their causal impact on predictions, which the authors argue is more accurate to how the model actually works.

The new method reveals more complete evidence for predictions: ViT-CX does a better job identifying all important visual evidence the model considers, rather than potentially missing key factors—important for understanding whether an AI system's reasoning is trustworthy.

Improved faithfulness means more reliable explanations: The paper reports "significantly better faithfulness to the model," meaning the explanations more accurately reflect what the model is actually doing internally, rather than providing misleading or incomplete accounts of its decision-making process.

Specialized explanation methods matter for different AI architectures: This work demonstrates that generic explanation techniques don't work equally well across different types of AI models—Vision Transformers require tailored approaches to produce trustworthy explanations.