The authors trained three versions of the same vision-language-action architecture, combining a vision-language backbone with a flow-matching action expert.
Reasoning supervision lets action models skip decoding without losing performance
TacitVLA matched explicit-reasoning policies across four robotics and driving benchmarks while reducing action latency by up to 3.4 times.
Big Tech
Jack Lu · Yurong You · Boyi Li · Yulong Cao · Thomas Tian · Chaowei Xiao · +5 more
NVIDIA · New York University · University of California, Berkeley · Johns Hopkins University · Stanford University
Research Digest··3 min read
Lu and colleagues test whether vision-language-action models must read their own newly generated reasoning before acting.
Why this paper
From NVIDIA and 4 others
In one line
Action generation does not need to wait for reasoning text; TacitVLA matches explicit-reasoning VLA performance at up to 3.4x lower latency.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks (5 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§