LatentLens Interactive Demo

About This Demo

This demo accompanies the paper "LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs". We train connector-only VLMs (frozen LLM + frozen vision encoder) and analyze interpretability of visual tokens at each layer using three methods:

• LatentLens (ours): Nearest neighbors in contextual text embeddings from a Visual Genome corpus
• LogitLens: Predicted tokens by applying the LM head to intermediate layer representations
• EmbeddingLens: Nearest neighbors in the LLM input embedding matrix

Click on any model cell below to explore per-patch interpretability across layers.