Editing the Goal: Per-Token SAE Steering and Its Depth Limits in $\pi_{0.5}$
Summary
- Researchers examine per-token sparse autoencoder (SAE) steering and depth limits in the robot foundation model pi_0.5.
- Replacing SAE features of a single goal word with an alternative goal word steers pi_0.5-LIBERO to complete alternative tasks.
- At layer 0, BatchTopK feature swaps successfully redirect behavior across tested scenes, while word-selected Matryoshka edits generally underperform.
- Adding a fixed difference-in-means vector at twice its natural magnitude also successfully redirects task rollouts.
- Layer-wise activation patching reveals that goal-word edits remain effective up to layer 6 for object location and layer 2 for object identity.
Why it matters
Understanding feature steering and depth limits in robot foundation models provides insights into internal goal representations and safety controls for physical AI systems.