Vision Transformers (ViTs) used in medical imaging are known to pick up spurious correlations—patterns that happen to align with labels but do not reflect genuine disease. Until now, these shortcuts have been difficult to trace inside the model, making them effectively impossible to correct. A new preprint on arXiv applies mechanistic interpretability to change that.
The authors examine a ViT-Base model trained for chest X-ray analysis and identify a specific attention head that encodes an annotation-driven bias. Rather than attributing the problem to the whole network or to the data alone, they localize the failure to a concrete internal component. This is a step toward understanding exactly how and where such biases emerge.
The finding matters because it suggests that spurious behavior in medical ViTs is not an opaque, distributed artifact but can be mechanistically isolated. If confirmed, this could enable targeted interventions—such as pruning or reweighting specific heads—rather than relying on broad retraining or dataset cleaning. The work is preliminary, and the abstract does not yet report the full experimental results or whether such fixes are effective in practice.