Mechanistic interpretability of VLA models
What do vision-language-action models represent, and how is their latent space organized?
Vision-Language-Action (VLA) models map images and instructions to robot actions, but we know little about how their internal representations are organized. I study the latent space geometry of VLA models with tools from mechanistic interpretability, such as sparse autoencoders and probing, to answer questions like:
- Which features drive an action decision, and are they grounded in the visual scene or the language instruction?
- How do vision, language, and action representations align, and where does that alignment break?
- Can collapse phenomena known from classification (Neural Collapse, minority collapse) explain failure modes of policies?
The goal is a geometric account of policy behavior that can be used to diagnose failures and shape representations during training.