TacitVLA hides reasoning from the action expert, cutting VLA latency up to 3.4x
A new paper introduces TacitVLA, a reasoning vision-language-action (VLA) model whose action expert never attends to the reasoning text generated during the same call, so actions do not have to wait for autoregressive decoding. The authors argue the reasoning supervision already improves the backbone representation, and across four benchmarks spanning manipulation and autonomous driving (RoboMME, RMBench, LIBERO and PhysicalAI-AV) they report matching an explicit-reasoning counterpart at up to 3.4x lower latency while matching the latency of the same architecture trained without reasoning at up to +17.1% success. In absolute terms the paper reports 82.7% on RoboMME Counting (vs. 66.8% prior best), 77.5% on RMBench (vs. 56.5%), 72.9% on LIBERO-Plus and a 52.5% scene score on the public split of the PhysicalAI-AV closed-loop driving challenge. All figures are self-reported and not yet independently reproduced.
If reproducible, this "tacit reasoning" approach could let robot-manipulation and driving policies keep reasoning gains while meeting real-time control latency budgets.