Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models

作者：Aiden Swann, Lachlain McGranahan, Hugo Buurmeijer, Monroe Kennedy III, Mac Schwager · 单位：Department of Mechanical Engineering, Department of Computer Science, Department of Aeronautics & Astronautics Stanford University · 会议/期刊：arXiv · 日期：2026-03-19 · 来源：Low-Level Learning-Based Action Modelling / Input Modelling / Evaluation of Vision Language Action Models