Reproduce “Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models” on OpenResearch

About this paper

Aiden Swann, Lachlain McGranahan, Hugo Buurmeijer, Monroe Kennedy III, Mac Schwager

Researchers at Stanford University developed a mechanistic interpretability pipeline using Sparse Autoencoders (SAEs) to extract interpretable, general, and steerable features from Vision-Language-Action (VLA) models. This approach identified features with 79.2% interpretability and enabled causal steering of robot behaviors, including those unachievable via language prompts, such as increasing gripper closure to 0.653 from 0.005.

Key insights

• SAEs can effectively extract highly interpretable features from VLA models, with 79.2% of sampled features aligning with human-identifiable concepts, significantly outperforming FFN neuron analysis.
• A small but critical fraction of SAE features encode general, reusable manipulation primitives that are causally essential for VLA performance, distinguishing them from task-specific memorizations.
• Directly steering these identified SAE features enables fine-grained control over robot actions, including the ability to elicit complex behaviors that are difficult or impossible to achieve through natural language prompts alone.

Summary by alphaXiv.