MOSAIC: sparse multimodal latent attribution for vision–language models (arXiv:2610.14190v1)
The paper introduces MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization), which learns a shared sparse image–text latent space with contiguous, phrase-aligned visual concepts and applies EK-FAC to decompose training-example influence through those concepts to specific phrases and image regions. The authors evaluate MOSAIC on Qwen and Gemma for captioning, chest X‑ray understanding, and visual question answering, reporting more localized and coherent features and that ablating influence-ranked latents alters outputs more than ablating latents ranked by activation or frequency.