TacitVLA trains on reasoning so actions need not wait for its generated text

When a model writes reasoning before acting, must the action wait for that text to benefit from reasoning training? The delay matters even on tasks where later actions never use the text.

The paper tests whether training on reasoning is enough, while keeping current-call reasoning out of the action’s inputs.

TacitVLA trains on reasoning text while keeping the current text out of action generation, so actions can start without waiting for it.

TacitVLA trains a model to generate reasoning text but masks that current-call text from its action expert. On memory tasks it still writes the next ledger entry, a running record used by later calls, while the current action executes.

Across manipulation and driving tests, its measured performance was close to Explicit Reasoning, whose action uses current-call text.

On LIBERO-Plus, it also scored above No Reasoning while generating no text at deployment, which the authors interpret as a benefit from reasoning supervision through the learned representation.

The results cover four tested task settings, and performance depends on the quality of the reasoning data. On memory tasks, generating a ledger entry during action execution avoids extra delay only when the entry finishes within that execution window; longer entries or shorter windows can add delay.