Noam Koenigstein

By aligning ViT attention with automatically generated, concept-level masks, this fine-tuning method substantially boosts robustness to distribution shifts, outperforming standard regularization techniques.