TacitVLA: NVIDIA's VLA drops reasoning latency 3.4x with no accuracy loss

机器之心 · wechat · 2026-11-08

Researchers from NVIDIA, NYU, UC Berkeley, JHU and Stanford propose TacitVLA: the model is trained to write reasoning text, but the action expert never reads the current step's reasoning — the benefit flows through training rather than inference-time input. Changing one attention mask suffices. Result: parity accuracy with explicit-reasoning VLAs across four robotics/driving benchmarks, up to 3.4x lower latency (147ms vs 495ms on LIBERO), and new SOTA on long-horizon memory benchmarks (RMBench 77.5% vs prior 56.5%).

Original post →

More from Embodied

Embodied channel →