Search papers, labs, and topics across Lattice.
This paper introduces Cross-Layer Transcoders (CLTs) as interpretable proxies for MLP blocks in Vision Transformers (ViTs), using an encoder-decoder scheme to reconstruct post-MLP activations from sparse embeddings of preceding layers. CLTs provide a layer-resolved decomposition of the final ViT representation, enabling attribution and process-level interpretability. Experiments on CLIP ViT-B/32 and ViT-B/16 show that CLTs achieve high reconstruction fidelity, preserve zero-shot classification accuracy, and reveal dominant layer-wise contributions.
Forget opaque embeddings: Cross-Layer Transcoders reveal how ViT layers contribute to the final representation, pinpointing the critical few that drive performance.
Understanding the internal activations of Vision Transformers (ViTs) is critical for building interpretable and trustworthy models. While Sparse Autoencoders (SAEs) have been used to extract human-interpretable features, they operate on individual layers and fail to capture the cross-layer computational structure of Transformers, as well as the relative significance of each layer in forming the last-layer representation. Alternatively, we introduce the adoption of Cross-Layer Transcoders (CLTs) as reliable, sparse, and depth-aware proxy models for MLP blocks in ViTs. CLTs use an encoder-decoder scheme to reconstruct each post-MLP activation from learned sparse embeddings of preceding layers, yielding a linear decomposition that transforms the final representation of ViTs from an opaque embedding into an additive, layer-resolved construction that enables faithful attribution and process-level interpretability. We train CLTs on CLIP ViT-B/32 and ViT-B/16 across CIFAR-100, COCO, and ImageNet-100. We show that CLTs achieve high reconstruction fidelity with post-MLP activations while preserving and even improving, in some cases, CLIP zero-shot classification accuracy. In terms of interpretability, we show that the cross-layer contribution scores provide faithful attribution, revealing that the final representation is concentrated in a smaller set of dominant layer-wise terms whose removal degrades performance and whose retention largely preserves it. These results showcase the significance of adopting CLTs as an alternative interpretable proxy of ViTs in the vision domain.
