NVIDIA's TacitVLA: robots reason without reading their own thoughts, 3.4x lower latency
arankomatsuzaki · x · 2026-10-16
NVIDIA and collaborators (NYU, UC Berkeley, Stanford, Johns Hopkins) propose TacitVLA, asking whether a VLA model's actions must condition on its current reasoning.
- The model still generates chain-of-thought or subtask reasoning, but the flow-matching action expert is trained to never read it.
- This keeps the gains of reasoning at up to 3.4x lower latency across 4 benchmarks, and tops RoboMME and RMBench.
- Implication: learning to generate reasoning is itself the benefit; actions need not attend to it step by step.
Related event: NVIDIA's TacitVLA Cuts VLA Latency 3.4x Without Losing Reasoning Gains(3 posts)→
More from Embodied
- WEDGE turns navigation agents' own failures into targeted training worlds — mohitban47 · 2026-10-16
- Knödel Watch: animated vocabulary cards on ESP32-S3, open-sourced 4 days after unboxing — alexcovo_eth · 2026-10-16
- SlipSense fuses pressure and vibration for 23ms zero-shot robotic slip detection — micoolcho · 2026-10-16
- Stanford's Fei-Fei Li team open-sources OpenWAM, a unified world-action model framework for robots — jiqizhixin · 2026-10-16
- Robot folding leaderboard: GPT-6 Astra leads at 1349 Elo, Muse Wonder trails at 1034 — qinzytech · 2026-10-16
- Google's Gemini 4 Argon aims to be the only AI model robots need, scores 87% on ERQA — The Decoder · 2026-10-16