Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies

This research introduces an event-grounded interpretability pipeline for Vision-Language-Action (VLA) policies. By anchoring sparse autoencoder (SAE) feature analysis to closed-loop behavioral events derived from end-effector trajectories, the method identifies features with strong causal effects on OpenVLA and transfers to continuous action chunks in $ ho_{0.5}$. Evaluated across two simulation architectures and a real-robot study, the approach demonstrates that SAE usability varies by architecture and intervention site, while aggressive interventions reveal safety and interpretability limits.