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

In three linesNovel mechanistic interpretability approach for Vision-Language-Action (VLA) robot policies. Authors propose sparse autoencoders (SAE) grounded in behavioral events rather than text contexts. Evaluation on OpenVLA and π₀.₅ across simulation and real-robot experiments, with code released.

Summary generated by Claude — human-verified