SpIn-ViT: A Sparsity-Induced Vision Transformer for Mechanistic Interpretability

A new framework named SpIn-ViT has been developed by researchers, which incorporates a sparsity-induced mechanism into Vision Transformers (ViTs) to improve mechanistic interpretability. In contrast to conventional post-hoc Sparse Autoencoders (SAEs) that rely on static representations, SpIn-ViT enables joint training of a pretrained ViT alongside a modified SAE in an end-to-end manner, ensuring that sparse patch-level representations align with the image classification goal. This method yields semantically meaningful neuron activations that pinpoint significant areas of images while still achieving competitive predictive accuracy. The framework underwent testing across nine image-classification benchmarks, utilizing classification accuracy, interpretability metrics, and evaluations from both AI and human reviewers. Results indicate that SpIn-ViT enhances the connection between learned features and downstream tasks, potentially improving the interpretability of vision models. The paper can be found on arXiv with the identifier 2608.14922.