Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders
For researchers using sparse autoencoders to interpret vision-language models, this work challenges a key assumption and provides a method to improve cross-modal alignment.
The paper identifies cross-modal feature heterogeneity in sparse autoencoders applied to vision-language joint embeddings, where the same concept activates different feature directions across image and text modalities. The proposed modality-specific sparse autoencoders with post-hoc alignment improve reconstruction fidelity and cross-modal retrieval performance.
Vision-language models map images and text into a joint embedding space. However, these embeddings often entangle multiple semantic features, which limits their interpretability and controllability. While sparse autoencoders have emerged as a useful tool for decomposing these embeddings into monosemantic features, their application to joint embedding spaces has largely relied on an implicit, untested assumption that semantically corresponding features share the same directions across modalities. In this paper, we challenge this assumption by identifying discrepancies in feature directions for the same concept across image and text modalities, a phenomenon we term cross-modal feature heterogeneity. We demonstrate that this heterogeneity is a key driver of the modality split, where a shared concept activates different latents depending on the modality. This finding further reveals why aligning latent activations alone is insufficient to resolve the underlying feature mismatch. Motivated by this observation, we propose an approach that trains modality-specific sparse autoencoders to preserve each modality's feature geometry, and then aligns corresponding features post hoc. Our method improves reconstruction fidelity and enhances performance in cross-modal retrieval and concept steering.