acceptodds
Under review as a conference paper at ICLR 2027

ReSync: Re-Aligning the Two Clocks of Asynchronous World-Action Models

Abstract

Jointly generating future video and actions has become a standard recipe for world-action models, and the strongest systems denoise the two streams on separate schedules: actions are decoded in few steps so control stays fast, while the video stream runs longer to keep the predicted future sharp. The design is deliberate, but it leaves the two streams on different clocks, and an action can become executable while the future that should justify it is still largely unresolved. We formalize this as a two-clock view of asynchronous inference and introduce the commitment-evidence gap, a quantity read directly from a model's own sampling schedule rather than measured by search. The gap is predictive: as it widens, candidate utility becomes harder to identify and extra candidate sampling buys less, while advancing the world stream buys more, and the two cross. Spending more world computation is therefore not simply better. The useful interval is closed at both ends, and both ends can be read off the schedule before any rollout. ReSync places the computation inside it: hold the action state, advance only the world within the supported window, then resume native denoising. No parameters change and no candidates are compared. On a frozen paired RoboCasa panel this improves success by 4.48 points, while an equal-compute control that waits without advancing the world does not move, and the same rule transfers to a second benchmark and a second backbone without retuning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world-action models, test-time compute, asynchronous denoising, diffusion policy, robot manipulation, inference-time scaling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vYk5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the "Commitment-Evidence Gap" in asynchronous world-action models for robotic manipulation, where low-latency action generation outpaces the slower prediction of future world states. To address this mismatch, the authors propose ReSync, an inference-time scheduling rule that briefly holds the action denoising process to allow world state denoising to advance, ensuring actions are conditioned on more mature future world predictions. The method is evaluated on RoboCasa, RoboTwin, Cosmos Policy, and a physical Mobile ALOHA setup.

Critical strengths
  1. Novel Problem Identification: The paper identifies a key timing mismatch between action and world clocks in multi-step generative models. This offers a valuable perspective on how inference timing affects decision quality.

  2. Plug-and-Play Applicability: ReSync is a pure inference-time scheduling adjustment requiring no model retraining. It demonstrates strong transferability across various datasets like RoboCasa and RoboTwin and architectures like Cosmos Policy.

Critical weaknesses
  1. High Inference Latency: The proposed holding mechanism seems to introduce a relatively high latency overhead. Considering the performance improvement is modest, this increased delay might limit its practical application in real-world robotic control.

  2. Presentation and Evaluation: The writing could be more accessible by avoiding the over-packaging of simple concepts with new terminology. Furthermore, including comparisons with other existing inference-time optimization methods would likely make the evaluation more persuasive.

Questions for authors

Clarification Requests:

  1. How did the physical robot handle the increased inference latency during the Mobile ALOHA experiments, and did this delay affect its ability to respond to dynamic changes?
  2. Would it be possible to include comparisons with other external inference-time optimization strategies to further strengthen the overall evaluation?

Suggestion for Future Work: 3. Have the authors considered integrating inference-time acceleration techniques to help offset the latency overhead introduced by the proposed method? This could be a valuable direction for future improvements.

Reviewer_i9oLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper looks at a timing mismatch in asynchronous world–action models: actions can become executable before the world model has had sufficient time to refine its prediction of the future to be of any use. It introduces the commitment–evidence gap as a schedule based proxy for this mismatch and shows empirically that as the gap increases, it becomes increasingly advantageous to advance the world prediction rather than to select among the available actions. The proposed method ReSync simply holds the action latent fixed, advances the world denoising and then resumes the original action schedule. The intervention is compatible with joint timestep support and leaves sufficient action updates to allow the refined prediction to be incorporated, requiring no retraining and no additional action denoising.

Validation is a combination of controlled schedule experiments, matched compute comparisons, placement and advancement ablations and transfer evaluations. On the paired RoboCasa rollouts ReSync achieves improved X-WAM success with increased forward compute, beating five other equal budget alternatives. The placement rule is supported by tests on two held out schedules and there are gains shown by RoboTwin and a constructed asynchronous Cosmos policy interface. A small physical robot study provides some initial feasibility evidence. The results in total support the notion of allocating extra compute to world prediction whilst actions can still adapt, with increased inference latency being the primary practical trade-off.

Critical strengths

The paper identifies a mismatch between action commitment and world refinement and then converts that into a very simple inference intervention which requires neither retraining nor any additional action updates. ReSync's placement rule takes into account both joint timestep support and the remaining opportunity for actions to adapt. This is significant as it provides an interpretable way of allocating test time computation, with experiments demonstrating when additional world refinement is of benefit and when that benefit starts to diminish.

On the paired RoboCasa evaluation ReSync achieves better success and beats the alternatives when run at the same forward budget. The placement and advancement ablations further probe the proposed mechanism and the fixed rule does well on two of the held out schedules. These comparisons go a long way to bolster the acceptance case as the gains are clearly attributable to the way and when the computation is applied rather than simply having more inference budget. The transfer results provide further evidence of applicability in different settings.

Critical weaknesses

The commitment-evidence gap is measuring schedule progress rather than decision relevant information and changing the world schedule also changes the quality of the prediction. The candidate selection and ReSync budgets are not matched in the experiment that is used to establish the claimed crossover. The separate matched budget comparisons do support the effectiveness of ReSync but they don't establish a general threshold at which evidence advancement becomes better than candidate selection. That in itself therefore limits the confidence you can place in the paper's broader explanation of how test time compute should be allocated.

The recovery margin is chosen empirically and the held-out schedule tests use the same X-WAM checkpoint and RoboCasa eval panel. The Cosmos experiment uses a deliberately constructed asynchronous interface with a fixed world advancement budget so it provides a more direct validation of the intervention transfer than transfer of the full schedule derived rule. Practical validation is also quite limited in that the primary configuration increases decision latency from 2.68s to 8.97s and the physical eval only contains 30 trials per condition. All of this leaves a fair amount of uncertainty as to whether the rule in question will actually lead to an improvement in the deployed system across native sampling schedules and in settings where timely replanning is required.

Questions for authors

Does the reported crossover between candidate selection and world advancement persist when their compute budgets are matched across schedules? The current crossover experiment uses unequal budgets; this distinction matters to the claim that the commitment–evidence gap predicts the preferable allocation of computation.

If addressed, I would consider increasing the score.

Reviewer_VunwOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies a timing mismatch in asynchronous world-action models (WAMs): actions are denoised in far fewer steps than future video, so the action commits before the world evidence that should justify it has resolved. The authors formalize a "commitment–evidence gap" G, show empirically across a controlled schedule family that the preferred use of test-time compute flips from candidate selection to world-evidence advancement as G grows, and propose ReSync — a parameter-free inference rule that inserts world-only denoising steps at a supported point where the action can still change, then resumes the native action schedule. On a frozen 960-rollout RoboCasa panel ReSync raises X-WAM from 80.31% to 84.79% at 3.40× forward compute, outperforming five equal-budget alternatives; gains transfer to RoboTwin 2.0 and a constructed two-clock Cosmos Policy interface.

Critical strengths
  • The two-clock framing and the G diagnostic are genuinely insightful and require no reward, critic, or extra model.

  • The equal-budget comparison (Table 2) is well designed: the same 34 shared-DiT forwards are spent on five natural alternatives, and ReSync beats all of them, which is far more convincing than reporting raw gains.

  • The interior optimum in the M-sweep (peak at M=24, decline at M=32/40) supports the claimed "usable window" mechanism and rules out a naive "more world compute is always better" explanation.

Critical weaknesses
  • Latency–performance tradeoff is under-discussed. The default operating point raises wall-clock latency from 2.68 s to 8.97 s per decision (0.37 Hz → 0.11 Hz, Table 11). The RoboCasa evaluator blocks on the policy response, so the practical closed-loop cost is substantial. The paper does not report the best achievable gain under a deployable frequency constraint (e.g., ≥0.2 Hz), where ReSync's advantage over cheaper baselines (warm-start, allocation) may shrink considerably.

  • Key comparison is marginally significant. ReSync vs. the strongest equal-budget baseline (world warm-start) is +1.87 pp with 95% CI [+0.09, +3.68] — the lower bound is essentially zero. With no multiple-comparison correction across the many contrasts in the paper, the headline claim that "where the compute lands matters more than how much" may not survive correction. The abstract-level wording should be softened.

  • Transfer claims are overstated. The Cosmos result is obtained on a constructed 14/35 two-clock interface, not the native Cosmos inference schedule. The core admissibility constraint (t_O ≥ t_a) also relies on X-WAM-specific asynchronous-noise training and the implementation detail ψ(j)=j. Generalization to other WAM families (autoregressive-video + IDM, fully decoupled streams, discrete action tokens) is not demonstrated.

  • Mechanism lacks direct measurement. The argument that inserted world steps produce "more resolved evidence" is supported only by a schedule-progress proxy. No direct measure of prediction quality (LPIPS, task-relevant feature consistency) is reported, and the reason for the decline beyond M=24 (training-support boundary vs. world drift vs. insufficient recovery steps) is not disentangled. The κ identifiability statistic is itself non-significant under the native schedule (p=0.34 on RoboCasa, negative on RoboTwin), weakening the diagnostic half of the story.

  • Real-robot evidence is underpowered (3 tasks × 10 resets; 19/30 vs. 22/30, not significant under exact McNemar). This is honestly framed as a feasibility check, but its presentation in the abstract risks over-reading.

Questions for authors

See the weakeness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.