NVIDIA TacitVLA Matches VLA Baselines at 3.4x Lower Latency
TL;DR
- TacitVLA matches explicit-reasoning VLA baselines at up to 3.4x lower latency across four manipulation and autonomous-driving benchmarks.
- On RMBench it reports 77.5% against a 56.5% prior best; on RoboMME Counting, 82.7% against 66.8%.
- The action expert is trained never to attend to the reasoning tokens the model generates in the current call.
TacitVLA, a vision-language-action model out of NVIDIA, keeps textual reasoning as a training signal but refuses to let the action head read any of those reasoning tokens at inference. The paper was posted this week.
The premise, in the abstract's own words: "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."
The authors, a group at NVIDIA with coauthors at Berkeley, Stanford and Johns Hopkins, run a controlled study across four benchmarks spanning manipulation and autonomous driving. TacitVLA, they write, "matches the performance of its explicit-reasoning counterpart at up to 3.4× lower latency, and matches the latency of the same architecture trained without reasoning at up to +17.1% success." On the RoboMME Counting suite it reports 82.7% against a 66.8% prior best; on RMBench, 77.5% versus 56.5%; and 52.5% on the public split of the PhysicalAI-AV closed-loop driving challenge.
For long-horizon tasks where the reasoning text doubles as episodic memory, the model writes each entry during action execution rather than before acting. That is the trick that keeps the latency win alive when memory matters.
Robotics is a dense beat right now; the paper is one of 163 robotics alerts in our tracker over the last 90 days.
Originally reported by huggingface.co
Read the original article →Original headline: NVIDIA's TacitVLA Matches Reasoning VLA Baselines With 3.4x Lower Latency by Decoupling Action From Text