Reasoning no longer blocks the action: masking current reasoning from the action expert keeps success and cuts latency up to 3.4x
Synopsis
Reasoning vision-language-action models previously made every action wait for autoregressive decoding of the reasoning text, which accounted for 53% to 70% of per-call latency. TacitVLA masks the action expert's attention to the current reasoning tokens, keeping only the training effect of reasoning supervision on the backbone representation, and on memory tasks generates each ledger entry during action execution. Across RoboMME, RMBench, LIBERO and PhysicalAI-AV it matches its explicit-reasoning counterpart while cutting latency up to 3.4x, and matches no-reasoning latency while improving success by up to 17.1%.
Interpretation
An action expert does not need to read the reasoning generated for the current call to obtain the full benefit of that reasoning. Embodied chain-of-thought made the action expert condition on the reasoning it had just generated, so every action chunk waited for that text to finish autoregressive decoding; on LIBERO and PhysicalAI-AV, generating the reasoning text accounts for 53% to 70% of per-call latency. On the RoboMME counting suite, TacitVLA scores 89.5, explicit reasoning 86.7 and no reasoning 79.0, against the benchmark's human reference of 88.5; on the RMBench four-task mean the three score 77.5, 77.0 and 60.4, all above WLA-0's 56.5.
On memory tasks, the reasoning entry that serves as episodic memory can be generated during action execution instead of before acting. Explicit reasoning runs generation and action in series, so the action waits for decoding to finish; TacitVLA computes the action first and generates the next ledger entry while the robot executes that chunk, and the entry enters the ledger at the same boundary under both policies. On the RoboMME counting tasks the action path takes 127 ms and generating the entry 93 ms, within 12% of the 800 ms execution window, so the action is ready after 127 ms versus 220 ms for explicit reasoning; on RMBench generation takes 66 to 120 ms, at most 1.1% of the 7.0 to 11.2 s window, and both policies keep a correct ledger in 99.7% of episodes.
On non-memory tasks, no reasoning text is generated at deployment, so latency equals the no-reasoning policy's. Explicit reasoning needs 495 ms before the action is ready on LIBERO and 269 ms in driving, spending that time generating reasoning text the next call never uses. TacitVLA takes 147 ms per call on LIBERO and LIBERO-Plus and 127 ms in closed-loop driving, the same as no reasoning and 3.4x and 2.1x lower than explicit reasoning; LIBERO's 250 ms window is met by TacitVLA and not by explicit reasoning.
Perspective
The report covers four benchmarks: RoboMME and RMBench as memory tasks, LIBERO and PhysicalAI-AV as non-memory tasks, with the three policies compared under the same sequence, data, prompts and optimizer while only the action expert's attention pattern changes. Latency is measured with per-stage timers on a single H100, and execution windows come from each simulator's own control rate. It suits researchers and engineers working on VLA inference cost and deployment scheduling.
Performance depends on the quality of the reasoning data, as for other reasoning VLAs. On memory tasks, deferring generation adds no latency only while it fits in the execution window, which it occupies at most 23% of in these benchmarks; longer entries or shorter windows would add the time by which generation exceeds the window. On non-memory tasks the reasoning text can still be generated during action execution for interpretability, a use the authors leave to future work.