- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies the effect of retaining computation state across observations in VLA action experts and introduces LycoReco, based on a Continuous Thought Machine. Paired carry/reset experiments examine the use of past information, internal update budgets, and action-chunk consistency. The complete model reports mean success rates of 97.5% on LIBERO, 64.6% on SIMPLER-WidowX, and 62.6% across five MIKASA tasks.
Critical strengths
The controlled experiments are the strongest part of the paper. Carry and reset models use matched architectures and training settings, allowing a fairly direct study of state lifetime. The paired-cue experiment in Table 2 is particularly effective: current observations are held identical while earlier cues change. Carry reaches the correct target first in 135 of 150 replays, whereas a deterministic memoryless policy is bounded at 50.
The compute-budget analysis provides evidence beyond tasks that explicitly require memory. With replanning at every step on LIBERO, carry with one internal update outperforms reset with twelve, and this comparison is repeated across three training seeds. Chunk disagreement and target-movement experiments further examine whether consistency is compatible with responding to new observations.
The paper is clearly written. Comparisons with published systems are distinguished from matched experiments on state lifetime, and the appendices provide detailed training settings, state definitions, and computational accounting.
Critical weaknesses
The efficiency conclusion currently concerns mainly the action expert's internal computation. Appendix C.2 reports complete-decision latency at K=1, but the invocation rates differ when K>1: carry processes observations within an action chunk, whereas some reset policies run once per chunk. Since the VLM accounts for most computation, fewer internal updates do not directly establish a proportional improvement during policy execution. Total computation and inference time over the same number of control steps would clarify the practical benefit.
Most comparisons still use one training seed. Paired scene intervals capture variation over evaluation scenes but not training variance. In addition, the CTM and recurrent Transformer in Appendix E differ in both parameter count and state capacity. The evidence for carrying state is stronger than the evidence for CTM's advantage over other recurrent architectures.
A smaller point concerns the target-movement experiment. Section 4.5 already states that the 100 scenes are selected from those originally solved by carry. The 65% and 100% results should therefore be interpreted as post-perturbation success on that subset, rather than recovery rates over the full task distribution. This supports responsiveness to new evidence, but not a broader claim of perturbation robustness.
Questions for authors
- At the same K and number of control steps, what are the total inference calls, computation, and inference time for carry and reset?
- Beyond the three-seed budget comparison, are repeated MIKASA training results or analyses of the CTM/Transformer state-capacity difference available?
3 comments
We additionally trained the following with the adapted encoder of the headline LIBERO system, while both VLM and action expert run once every steps:
LIBERO:
| Method | ||
|---|---|---|
| 1-tick carry | 97.00% | 97.25% |
| 12-tick carry | 97.50% | 97.50% |
| 1-tick reset | 90.25% | 97.25% |
| 12-tick reset | 90.75% | 97.00% |
While both methods saturate near ~97% at , carry maintains this performance even under every-step replanning (), where reset drops by roughly 7 percentage points.
Practical implications under latency constraints. For latency-constrained robotic control, achieving stronger policy performance within the available inference budget is important. Model capacity and tick count offer two complementary scaling dimensions: larger experts provide greater representational capacity, while additional ticks enable further iterative refinement. Carrying state can improve this tradeoff rather than eliminate the need for either. In our LIBERO budget sweep, carry outperforms reset at every matched tick budget, while additional ticks can still improve carry's performance. Reducing the required tick budget may also create headroom for larger experts within the same latency constraint. While 1-tick carry does not universally surpass 12-tick reset across every evaluation regime (such as the saturated setting above), carry consistently achieves higher performance at every matched tick budget.
Potential benefits in dual-system VLAs. When an action expert operates more frequently than the VLM, its inference latency becomes a key constraint on control frequency. Our dual-rate experiment (Section 4.6) shows that carry maintains its advantage over reset when the VLM refreshes less frequently, supporting its applicability to such architectures. At 's expert scale (roughly 100 times larger), a hypothetical LycoReco-like expert would require approximately 0.86 GFLOPs at one tick versus 10.3 GFLOPs at twelve (Table 6). However, this larger model has not been trained due to compute budget. We will clarify these limitations and distinguish demonstrated expert-level efficiency from end-to-end benefits.
Most comparisons still use one training seed
We have expanded our evaluations across three independent training seeds per benchmark in the next revision. Success rates (%) are reported as mean ± sample standard deviation. The budget sweep results are provided in our response on practical efficiency above.
We also provide the headline benchmark results:
| Model | LIBERO | SIMPLER-WidowX | MIKASA (-VLA-v0 release) | MIKASA (-v0 release) |
|---|---|---|---|---|
| LycoReco | 97.25 ± 0.25 | 65.16 ± 5.62 | 59.33 ± 5.66 | 60.50 ± 4.10 |
In addition, the CTM and recurrent Transformer in Appendix E differ in both parameter count and state capacity
To isolate architectural differences from state capacity in Appendix E, we trained a Recurrent Transformer (RT) control matched to CTM's exact retained-state size (82 KiB) and trainable parameter count (2,996,301 for RT vs. 2,995,790 for CTM) using the same frozen PaliGemma backbone and training recipe on MIKASA. The per-task success rates (%) are reported below:
| Expert | Lifetime | SGT | IM | RC3 | RC5 | RC9 | Mean |
|---|---|---|---|---|---|---|---|
| LycoReco | carry | 61 | 24 | 89 | 52 | 12 | 47.6 |
| Matched Recurrent Transformer | carry | 45 | 25 | 63 | 31 | 10 | 34.8 |
A smaller point concerns the target-movement experiment
We agree. This experiment demonstrates responsiveness to new visual evidence within the scenes solved without perturbation. We do not claim general perturbation robustness, and we will clarify this distinction in Section 4.5 and the abstract/introduction to avoid the ambiguity raised.
Thank you to the authors for their point-by-point response to the review comments. The supplementary experiments and clarifications are reasonably thorough and have addressed most of my concerns. However, some issues remain only partially resolved, so I will maintain my score of Weak Accept.
We sincerely thank the reviewer for their thoughtful feedback and valuable suggestions. Below, we provide the requested efficiency measurements, three-seed evaluations, and matched-capacity architectural comparisons.
We agree that practical efficiency should be compared over the same number of control steps, counting both VLM and action-expert computation.
Carrying state does not require processing every observation within an action chunk: state lifetime and inference frequency are independent design choices. A carry expert can keep its state across chunk-level decisions without updating it at intermediate control steps, and our dual-rate experiment (Section 4.6) likewise runs the VLM less often than the expert. Appendix C.2 described only the every-step schedule, and we will revise it. We compared carry and reset under a matched schedule in which both the VLM and the expert run once every K steps (experimental setup from Section 4.4). Over the same control steps, both make the same number of VLM and expert calls at identical total computation:
We acknowledge that end-to-end savings are modest in our implementation. With a ~3M-parameter expert and a ~3B-parameter VLM, the VLM dominates total computation, so reducing expert ticks does not yield proportional full-policy latency savings. Nevertheless, our budget experiment (Section 4.4) demonstrates a favorable performance–computation tradeoff: carry outperforms reset at every matched tick budget on the frozen PaliGemma. Reducing ticks from T=12 to T=1 lowers action-expert latency from 5.835 ms to 1.279 ms.