TacitVLA: Reasoning Vision-Language-Action Models Without Reasoning Latency
Abstract
Textual reasoning before acting has improved vision-language-action (VLA) models, but at the cost of latency: every action conditioned on the reasoning text must wait for autoregressive decoding. We show that actions do not need to wait for or condition on the generated reasoning to benefit from it, because the reasoning supervision already improves the backbone representation. We introduce TacitVLA, a reasoning VLA whose action expert never attends to the reasoning it generates for the current call. On long-horizon tasks where the reasoning text also serves as episodic memory, it can generate each entry during action execution rather than before acting. We evaluate it in a controlled study on four benchmarks spanning manipulation and autonomous driving, memory and non-memory tasks: RoboMME, RMBench, LIBERO, and PhysicalAI-AV. On all four, TacitVLA matches the performance of its explicit-reasoning counterpart at up to 3.4x lower latency, and matches the latency of the same architecture trained without reasoning at up to +17.1% success. In absolute terms, it surpasses the best previously published results on the RoboMME Counting suite (82.7% vs. 66.8%) and RMBench (77.5% vs. 56.5%), and reaches 72.9% on LIBERO-Plus and a 52.5% scene score on the public split of the PhysicalAI-AV closed-loop driving challenge. Project page: https://tacitvla.github.io.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
TacitVLA demonstrates that reasoning supervision improves VLA performance through representation learning rather than explicit inference-time conditioning, enabling real-time deployment with up to 3.4x lower latency. This paper provides a rigorous, controlled study across manipulation and driving benchmarks, showing that masking the action expert's attention to reasoning tokens retains the performance benefits of reasoning while eliminating the autoregressive decoding bottleneck, a significant contribution to the efficiency of embodied AI systems.
Methodology
The paper proposes TacitVLA, a method that decouples the *generation* of reasoning text from the *conditioning* of actions on that text. The core insight is that the benefit of reasoning in Vision-Language-Action (VLA) models primarily stems from the representation learning induced by the reasoning supervision signal (cross-entropy loss on text tokens) during training, rather than the explicit attention to the generated text tokens during inference. By masking the action expert's attention to the reasoning tokens during both training and inference, the model retains the improved backbone representations without incurring the autoregressive decoding latency. For memory tasks, the paper introduces an asynchronous generation strategy where the reasoning text (ledger entry) is generated during the execution of the current action chunk, ensuring it is available for the next call without blocking the current action. This is a clever architectural optimization that leverages the existing Mixture-of-Transformers structure.
Results & evaluation
The evaluation is rigorous and comprehensive, covering four distinct benchmarks: RoboMME (memory/counting), RMBench (long-horizon manipulation), LIBERO/LIBERO-Plus (standard manipulation/robustness), and PhysicalAI-AV (autonomous driving). The controlled study comparing Explicit Reasoning, TacitVLA, and No Reasoning is well-designed, isolating the effect of attention masking. The results show that TacitVLA matches the performance of Explicit Reasoning while achieving the latency of the No Reasoning baseline (or slightly higher due to deferred generation on memory tasks). The latency improvements are significant (up to 3.4x), which is critical for real-time deployment. The performance gains over the No Reasoning baseline (up to +17.1% success) confirm the value of the reasoning supervision.
Reproducibility
The paper provides detailed implementation details in the appendix, including architecture specifics, training hyperparameters, data sources, and evaluation protocols. The use of standard benchmarks (LIBERO, RoboMME, RMBench) and public datasets (PhysicalAI-AV public split) enhances reproducibility. The code and project page are available. However, the driving dataset is proprietary (3.8M events), which limits full reproducibility for that specific benchmark, though the methodology is clear.
Limitations
The primary limitation is the reliance on high-quality reasoning annotations for training. If the reasoning data is noisy or incorrect, the representation learning benefit may diminish. Additionally, the asynchronous generation on memory tasks assumes that the generation time fits within the action execution window; if the reasoning text becomes very long or the hardware is slower, this assumption may break, leading to latency spikes. The paper also notes that on non-memory tasks, the reasoning text is discarded, which might miss opportunities for interpretability or safety checks that could be performed post-hoc.
Broader impact
This work has significant implications for the deployment of VLA models in real-world robotics and autonomous driving, where latency is a critical constraint. By demonstrating that reasoning can be "internalized" into the model's representations without the inference cost, it paves the way for more efficient and responsive embodied AI systems. The approach is generalizable to other architectures that use reasoning as a training target, potentially influencing future designs of foundation models for robotics. TacitVLA demonstrates that reasoning supervision improves VLA performance through representation learning rather than explicit inference-time conditioning, enabling real-time deployment with up to 3.4x lower latency. This paper provides a rigorous, controlled study across manipulation and driving benchmarks, showing that masking the action expert's attention to reasoning tokens retains the performance benefits of reasoning while eliminating the autoregressive decoding bottleneck, a significant contribution to the efficiency of embodied AI systems.
Author of this paper? Send a correction.