Token-Based Detection of Spurious Correlations in Vision Transformers

A new research paper on arXiv (2509.04009) presents a token-based diagnostic pipeline for detecting spurious correlations in Vision Transformers. Spurious correlations occur when models rely on unintended patterns, such as color aberrations or text fragments, that coincide with the task but are not causally linked. The proposed method applies a leave-one-out approach to identify tokens that contribute to predictions based on spurious features. This work addresses the critical need for trustworthy and generalizable machine learning models in computer vision.