NVIDIA Researchers Unveil TacitVLA, Boosting Robot Action Speed by 3.4x Without Reasoning Latency

The first author of this paper is Jack Lu, a PhD student in Computer Science at New York University, and the research work was completed during his internship at NVIDIA. This paper was co-supervised by Yan Wang, a Senior Research Scientist at NVIDIA and Head of Alpamayo Research Technology, and Marco Pavone, Head of Autonomous Driving Research at NVIDIA and Associate Professor at Stanford University.

Nowadays, many Vision-Language-Action (VLA) models reason before acting. Before outputting an action, the model first writes a short piece of text, such as a description of the scene, a plan, or a sub-task to be completed next. Doing this results in higher success rates for robots, but also slower speeds. Reasoning text needs to be generated autoregressively token by token, and actions must wait for the text to be finished before they can start. In the experiments of the paper, writing reasoning text accounted for up to 70% of the time of a single model call.

A research team from NVIDIA, New York University, UC Berkeley, Johns Hopkins University, and Stanford University carefully examined where the benefits of reasoning actually come from. They found that actions do not need to read the reasoning written in the current step; the model learning to write reasoning is already sufficient. Based on this, the team proposed TacitVLA. Like standard reasoning VLA, it is trained to write reasoning, but its actions never wait for this text.

On four benchmarks of robot manipulation and autonomous driving, TacitVLA's success rate is comparable to explicit reasoning, with latency reduced by up to 3.4 times. At the same speed, it is up to 17 percentage points higher in success rate than the same model without reasoning training. On two long-term memory benchmarks, TacitVLA also refreshed the best scores of published methods, improving by 21 percentage points on RMBench and by 16 percentage points on the counting task of RoboMME.

Paper Title: TacitVLA: Reasoning Vision-Language-Action Models Without Reasoning Latency

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

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

Three functions of the reasoning text

In reasoning VLA, the generated text has three functions.

During training, it is a supervision target. Learning to write this text will shape the backbone that the action expert relies on.

In long-horizon tasks, it can also serve as memory. Suppose a robot needs to press a button several times; the camera cannot capture how many times it has already been pressed, but a log of completed sub-tasks can.

Finally, during deployment, the text is the input of the action expert. The action expert needs to read the text written in the current step, so the action must wait. TacitVLA retains the first two roles and removes the third.

Most recent reasoning VLAs adopt the same structure. The vision-language backbone is responsible for writing reasoning, and the action expert based on flow matching reads the backbone's information through attention and outputs actions. On each benchmark, the team used this structure to train three policies on the same data.

In Explicit Reasoning, the action expert reads the reasoning it just wrote, consistent with the approach of embodied chain-of-thought. TacitVLA uses the exact same training sequence, loss function, prompt, and optimizer; the only difference is a change in the attention mask, which prevents the action expert from seeing the reasoning tokens. No Reasoning is a model trained on the same structure on data without reasoning text.

Such a design separates the two pathways where reasoning might provide help.

First, learning to write reasoning may train a better backbone. Second, reading reasoning may provide useful additional input for the action expert. Explicit Reasoning possesses both of these effects, while No Reasoning has neither. TacitVLA only retains the first type, and the first type does not consume any time during deployment.

During deployment, if the reasoning text is of no use, TacitVLA simply skips writing it. In memory tasks, each new sub-task still needs to be written to the log for use in subsequent steps. Therefore, TacitVLA first calculates the action, and then writes the log while the robot is executing this action. On the two memory benchmarks, the log is written before the action execution is finished in every call, so it does not bring any additional delay.

Experimental results

On all four benchmarks, TacitVLA's success rate is comparable to Explicit Reasoning and exceeds No Reasoning, while maintaining the speed of No Reasoning.

The greatest improvement is in the memory benchmark. On RMBench, TacitVLA achieved an average success rate of 77.5% across four tasks, Explicit Reasoning was 77.0%, and the previously published best result was 56.5%. Its improvement relative to No Reasoning (60.4%) mostly comes from the Press Button task, in which only the number of presses is recorded in the logs. On this task, the success rate of No Reasoning was 7.7%, and TacitVLA was 74.3%.

On RoboMME, when trained on all 16 tasks, TacitVLA's overall success rate was 48.1%, with the previous best result being 44.5%. On counting tasks, it reached 82.7%, with the previous best being 66.8%. When trained only on counting tasks, it reached 89.5%, higher than the human reference score of 88.5% provided by the benchmark. In the human reference setting, humans are only responsible for selecting high-level steps, while specific actions are executed by an oracle planner, whereas every action of TacitVLA is completed by itself.

Reasoning is also helpful on tasks that do not require memory. LIBERO-Plus tests seven categories of perturbations in a zero-shot manner. On this benchmark, the two reasoning strategies achieved the highest scores among published strategies trained only on LIBERO, with Explicit Reasoning at 73.5%, TacitVLA at 72.9%, and No Reasoning at 67.6%.

On the standard LIBERO, TacitVLA reaches 98.6%, tying with the best published results. When performing closed-loop driving tests on the public AlpaSim challenge scenarios, TacitVLA's scenario score is 52.5%, Explicit Reasoning is 52.3%, and No Reasoning is 46.4%. This performance is only 2.5 points away from Alpamayo-R1. Alpamayo-R1 is a 10B model using four camera heads, while TacitVLA's number of parameters is less than half of it, and it also uses only half the cameras.

On a single H100 GPU, TacitVLA requires 147 ms to output an action on LIBERO, while Explicit Reasoning requires 495 ms, which is 3.4 times the former. In autonomous driving, the two are 127 ms and 269 ms, respectively. On memory tasks, Explicit Reasoning is 1.4 to 1.7 times slower. TacitVLA, on the other hand, writes logs during execution, occupying at most 23% of the execution time.

Summary

The experiments covered robot manipulation and autonomous driving, and also covered tasks that require memory and tasks that do not require memory. In these settings, the benefits of reasoning are passed to actions through training on reasoning text. On memory tasks, the benefits are also passed through previously written logs. Reading the reasoning of the current step only makes the policy slower and does not make it more successful.

TacitVLA does not need additional models, nor does it need a scheduler. Any reasoning VLA adopting this structure can become a TacitVLA as long as the action expert is prevented from seeing the reasoning tokens using an attention mask during training.

© THE END

For republication, please contact this official account for authorization

For submissions or media inquiries: [email protected]