●builderYou can deploy reasoning-capable VLA models with much lower latency by removing the need for action conditioning on reasoning text.
●researcherThis demonstrates that reasoning supervision can improve representations even when the action head does not explicitly attend to the reasoning tokens.