TacitVLA: NVIDIA's Reasoning-Free VLA Cuts Robot Action Latency 3.4x

Researchers from NVIDIA, NYU, UC Berkeley, Johns Hopkins, and Stanford introduce TacitVLA, a vision-language-action (VLA) model that retains the training benefits of explicit reasoning while removing its inference latency. In standard reasoning VLAs, the model first generates reasoning text (scene descriptions, plans, or sub-tasks) autoregressively before outputting actions, causing reasoning to consume up to 70% of per-call latency.

Three Roles of Reasoning Text

The authors identify three roles of generated reasoning text in VLAs:

Training supervision: Learning to produce reasoning shapes the backbone representations used by the action expert.

Long-horizon memory: Reasoning logs record completed sub-tasks (e.g., button-press counts) when visual observation cannot.

Deployment input: The action expert attends to the current step's reasoning tokens, forcing actions to wait for text generation.

TacitVLA preserves the first two roles and discards the third.

Method

All compared policies share the same architecture: a vision-language backbone that writes reasoning, and a flow-matching action expert that reads backbone features via attention. Three variants are trained on identical data, loss, prompts, and optimizer:

Explicit Reasoning: Action expert attends to reasoning tokens (standard embodied chain-of-thought).

TacitVLA: Identical training sequence, but an attention-mask change prevents the action expert from seeing reasoning tokens.

No Reasoning: Trained on data without reasoning text.

At deployment, TacitVLA skips reasoning generation entirely when not needed for memory. On memory tasks, it computes actions first and writes logs in parallel during execution; logs finish before the action completes, adding zero extra latency.

Experimental Results

Evaluated on four benchmarks spanning robot manipulation and autonomous driving:

RMBench (long-horizon memory): TacitVLA averages 77.5% success vs. 77.0% for Explicit Reasoning and 56.5% prior best. On the Press Button task (counting presses only via log), No Reasoning scores 7.7% while TacitVLA reaches 74.3%.

RoboMME (16 tasks, includes counting): TacitVLA overall 48.1% (prior best 44.5%). Counting task: 82.7% (prior best 66.8%). Trained only on counting tasks, TacitVLA hits 89.5%, exceeding the human reference of 88.5% (where humans select high-level steps and an oracle planner executes actions).

LIBERO-Plus (zero-shot perturbations): Explicit Reasoning 73.5%, TacitVLA 72.9%, No Reasoning 67.6% — highest among LIBERO-trained policies.

LIBERO (standard): TacitVLA 98.6%, matching prior best.

AlpaSim closed-loop driving: TacitVLA scene score 52.5%, Explicit Reasoning 52.3%, No Reasoning 46.4%. Alpamayo-R1 (10B params, 4 cameras) scores 55.0%; TacitVLA uses less than half the parameters and half the cameras.

Latency Measurements

On a single H100 GPU:

LIBERO: TacitVLA 147 ms/action vs. Explicit Reasoning 495 ms/action (3.4× slower).

Autonomous driving: TacitVLA 127 ms vs. Explicit Reasoning 269 ms (2.1× slower).

Memory tasks: Explicit Reasoning 1.4–1.7× slower; TacitVLA's logging occupies at most 23% of execution time.

Conclusion

Across manipulation and driving, with and without memory demands, the benefit of reasoning transfers to actions through training-time supervision and, for memory tasks, through previously written logs. Reading the current step's reasoning only adds latency without improving success. TacitVLA requires no extra models or schedulers; any reasoning VLA with this architecture can adopt it by masking reasoning tokens from the action expert's attention.

Paper: https://arxiv.org/abs/2610.17540

Project page: https://tacitvla.github.io/