NVIDIA Researchers Unveil TacitVLA, Boosting Robot Action Speed by 3.4x Without Reasoning Latency
Overview
Researchers from NVIDIA, NYU, UC Berkeley, Johns Hopkins, and Stanford present TacitVLA, a Vision-Language-Action model that eliminates the latency of autoregressive text generation during action execution. By modifying an attention mask to prevent the action expert from reading current reasoning tokens, the method retains training benefits while skipping inference delays. Experiments on robot manipulation and autonomous driving benchmarks show comparable success rates to explicit reasoning models but with up to 3.4x faster action output. On memory-intensive tasks like RMBench, TacitVLA matches or exceeds explicit reasoning performance while maintaining non-reasoning model speeds.
Generated from attributed reports · Updated 2 days ago
Event evidence and corrections
1 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · Reported success rate: 77.5 percent · reported trials · Differing source assertions
“在 RMBench 上，TacitVLA 在四个任务上的平均成功率为 77.5%，Explicit Reasoning 为 77.0%，此前已发表的最好结果为 56.5%。”
Exact source · revision 1jiqizhixin
Reported quantity · Reported success rate: 7.7 percent · reported trials · Differing source assertions
“在该任务上，No Reasoning 的成功率为 7.7%，TacitVLA 为 74.3%。”
Exact source · revision 1jiqizhixin
Reported quantity · Reported success rate: 48.1 percent · reported trials · Differing source assertions
“在 RoboMME 上，用全部 16 个任务训练时，TacitVLA 的总体成功率为 48.1%，此前最好结果为 44.5%。”
Exact source · revision 1jiqizhixin
Report timeline
Follow attributed reports and material updates.
- Machine Heart — Robotics on WeChatSignalNVIDIA Researchers Unveil TacitVLA, Boosting Robot Action Speed by 3.4x Without Reasoning Latency
Researchers from NVIDIA, NYU, UC Berkeley, Johns Hopkins, and Stanford present TacitVLA, a Vision-Language-Action model that eliminates the latency of autoregressive text generation during action execution. By modifying an attention mask to prevent the action expert from reading current reasoning tokens, the method retains training benefits while skipping inference delays. Experiments on robot manipulation and autonomous driving benchmarks show comparable success rates to explicit reasoning models but with up to 3.4x faster action output. On memory-intensive tasks like RMBench, TacitVLA matches or exceeds explicit reasoning performance while maintaining non-reasoning model speeds.
Event coverage history
Current attention 3·Peak within the comparable range 10（2026-11-08 05:00 UTC）·Change within the comparable range over 24 hours -51%
The trend compares participants observed continuously; its scope may be narrower than current attention totals. Hover or click to explore hourly observations, or use the left and right arrow keys.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).