A new paper on arXiv challenges a widespread assumption in computer vision: that vision transformer (ViT) classifiers can be meaningfully explained by decoding individual patch tokens. This technique has become a standard diagnostic tool, largely because it tends to produce seemingly intuitive visual explanations.
The paper, titled "The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability?" (arXiv:2610.25289), directly questions whether this behavior is a genuine property of the models or a kind of illusion. The abstract notes that the assumption is widely held, but the title suggests the authors argue it is not necessarily true.
Given the abstract is truncated, the full details of their argument are not yet available. However, the framing is clear: researchers should not take patch-decoding interpretability for granted in vision transformers, and the paper urges a closer look at what these explanations actually represent.