LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs

LatentLens is a novel interpretability method designed to enhance the understanding of visual tokens within large language models (LLMs) when transformed into vision-language models (VLMs). It utilizes a shallow MLP to map visual tokens into the LLM's embedding space and compares these to contextualized token representations from a large text corpus, yielding semantically meaningful descriptions that improve interpretability across 15 VLMs. This advancement is significant for practitioners as it provides deeper insights into the alignment of vision and language representations, facilitating better analysis of latent representations in AI models.