We introduce Token Steering: an inference-time method for steering pretrained autoregressive VLAs through direct intervention in their action-token space—no retraining or fine-tuning required.
Autoregressive VLAs generate actions token-by-token. We show that this representation can also be used as an interface for real-time human guidance. The policy retains its learned dexterity while users correct mistakes, resolve ambiguity, and recover from failures.
Across two household manipulation tasks, Token Steering improves success rates from 10.0% -> 72.5% and 16.7% -> 93.8%.
Paper: arxiv.org/abs/2606.15021

