acceptodds
Under review as a conference paper at ICLR 2027

InstantWAM: Instant World Action Model via Asymmetric One-Step Decoding

Abstract

World Action Models (WAMs) regenerate every action plan from noise, even though much of the next plan was already predicted at the previous replan. This symmetry wastes computation and ignores the temporal structure of closed-loop control. We introduce InstantWAM, built on a simple asymmetry: generate the first plan, then update the future already in motion. InstantWAM carries the unexecuted suffix into the next planning window as a persistent generative state, perturbs it back onto the pretrained flow, and corrects it with a single observation-conditioned evaluation. The WAM's own action attention determines how long the inherited future remains useful. Repeated multi-step generation thereby becomes one-step decoding without training or changing the frozen model. InstantWAM reduces average network function evaluations from 10 to 1.62, delivering faster planner calls and faster episode-level planning. On RoboTwin, it achieves success versus for 10-step FastWAM; on LIBERO, it reaches . The future should not be discarded at every replan, but carried forward and updated.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, World Action Models, Generative Action Policies, Flow Matching, Inference Acceleration

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_DuJ3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes InstantWAM, a training-free inference scheme for flow-matching World Action Models (WAMs), instantiated on FastWAM. After a standard first plan generated with 10 network calls, each subsequent replan works as follows: (i) shift the unexecuted part of the previous action chunk and pad it by repeating the last action; (ii) re-noise it to the last nonzero timestep of FastWAM’s sampling schedule; (iii) correct it with a single observation-conditioned network call. The paper reports reducing the average number of network calls per plan from 10 to 1.62, giving 3.19× faster planner calls and 2.47× faster episode-level planning, achieving comparable success rates to FastWAM w/o inference speedup on RoboTwin 2.0 and LIBERO.

Critical strengths
  • S1. Simple and easy to adopt. The method is training-free and requires no modification to the frozen model.
  • S2. Substantial end-to-end speedup. The speedup is measured as end-to-end planning latency, not only as a reduction in network calls.
Critical weaknesses

W1. Limited novelty and missing related work. Reusing the unexecuted part of a previous action chunk to accelerate the next generation has already been proposed in training-free form for pre-trained generative policies, e.g., Falcon (ICML 2025) and RTI-DP (IROS 2025). These works are not cited. The authors should discuss them in the related work and clearly state what is novel beyond transferring the idea to FastWAM’s flow-matching decoder, especially relative to RTI-DP, which appears highly similar (RTI-DP also shifts the unexecuted part of the previous chunk, repeats the last action, and generates a new action chunk from it). Candidate differences could be re-noising for the reused action chunk, the single network call update. These should be articulated and their advantages should be supported empirically.

W2. Insufficient experimental controls. (i) Table 1 and Table 5 should include a call-matched baseline. Although Table 2 includes Fresh Noise (1 call from noise per plan), its first plan also uses only 1 call, so its budget differs from InstantWAM’s (10 calls for the first plan, then 1 call), and on RoboTwin it is evaluated only on a 15-task subset. A Fresh Noise variant with the same 10→(1 or 2) schedule, evaluated on the full benchmarks, would isolate how much reusing the previous chunk helps under an identical computational budget.

(ii) Table 1 and Table 5 report point estimates only. Please state the number of random seeds, report standard deviations, and indicate in the main text which benchmark and episodes Table 5 is measured on.

(iii) All experiments use FastWAM only. Results on other WAM backbones would further demonstrate the generality of the method.

[1] Chen et al. Falcon: Fast Visuomotor Policies via Partial Denoising. ICML, 2025.

[2] Duan et al. Real-Time Iteration Scheme for Diffusion Policy. IROS, 2025.

Questions for authors

Q. What exactly does Direct Prior compute at each replan? Two interpretations seem possible to me: (a) the unexecuted and padded action chunk is executed without a policy call, or (b) the unexecuted and padded action chunk is fed to the policy without re-noising. Sec. 5.3 says "Direct Prior cannot absorb the new observation", which supports (a). However, Table 8 reports a 10→1 schedule and a 144.0 ms update latency (equal to one call), which supports (b). If (b), the observation is actually used.

Reviewer_r2LiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The authors propose a training-free inference acceleration method for world action models with action chunking. Since only a portion of the generated action trajectory is executed before replanning, the authors propose reusing the unexecuted portion of the previous plan. Specifically, the method shifts the remaining actions and mixes them with Gaussian noise to construct an intermediate flow state, from which a new action trajectory is generated using a single observation-conditioned velocity evaluation. The authors provide a conditional error analysis and evaluate the method on RoboTwin 2.0 and LIBERO, reporting near-baseline success rates with substantially reduced planning latency.

Critical strengths

The proposed method substantially reduces planning latency while maintaining near-baseline success rates. Although the computational reduction primarily comes from replacing multi-step inference with one-step updates, maintaining competitive task performance without additional training is practically valuable.

Critical weaknesses

The pretrained velocity field is conditioned on both the flow timestep and the current observation, and is trained on the corresponding intermediate distributions. However, the proposed method constructs its intermediate state by mixing a shifted action trajectory from the previous observation with Gaussian noise and evaluates the velocity field at a fixed intermediate timestep under the new observation. This constructed state does not necessarily follow the intermediate distribution associated with the specified timestep and current observation. Therefore, it remains unclear whether the learned velocity field can provide an accurate correction from this potentially mismatched state, particularly with only a single Euler update.

Although the authors provide a conditional error bound, its key quantities are not fully evaluated, and the empirical diagnostics do not establish the uniform contraction conditions required for the asymptotic guarantee. Furthermore, showing that the shifted prior is closer to the teacher's intermediate state than Gaussian noise is largely expected, given that the prior already contains information about the previously generated action trajectory. This result alone does not sufficiently validate the accuracy of the proposed one-step correction.

The central idea of reusing previous action predictions to reduce inference steps is closely related to RTI-DP (Duan et al., IROS 2025), which already proposes full initialization followed by shifted-plan refinement without retraining. RTI-DP also discusses potential extensions to flow-based models and one-step inference. Although InstantWAM introduces a specific flow re-entry formulation and attention-guided replanning, its methodological novelty relative to this prior work is not sufficiently clarified or experimentally established.

Although the authors compare against one-step FastWAM through the Fresh Noise baseline, the comparison does not fully isolate the benefit of temporal reuse. Fresh Noise uses one NFE from the initial plan, whereas InstantWAM uses a ten-NFE bootstrap followed by one-step updates. A comparison using the same initial bootstrap and subsequent inference budget would more clearly demonstrate the contribution of temporal reuse.

Although attention-guided replanning improves performance over fixed H=24, it achieves only a marginal improvement over fixed H=18, which is close to its average selected horizon. Therefore, it remains unclear whether adaptive horizon selection provides a meaningful advantage over a well-chosen fixed execution horizon.

Questions for authors
  • How does the proposed method justify applying the pretrained time- and observation-conditioned velocity field to an intermediate state constructed from the previous action trajectory? Under what conditions can a single Euler update from this potentially mismatched state accurately approximate the target action distribution?
  • Can the authors quantify the key terms in the proposed error bound and evaluate its tightness empirically? In particular, is there evidence that the contraction conditions hold in practice?
  • How does the proposed method differ from RTI-DP (Duan et al., IROS 2025), which also reuses previous action predictions for faster inference? What are the specific advantages of the proposed flow re-entry mechanism over this prior approach?
  • Could the authors compare against a one-step Fresh Noise baseline using the same ten-NFE initial bootstrap? This would help isolate the contribution of temporal reuse from initialization quality.
  • Although the authors argue that attention enables state-dependent horizon selection, its performance is comparable to fixed H=18. What practical advantage does adaptive selection provide over a well-chosen fixed horizon?
Reviewer_xh79OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

InstantWAM accelerates frozen world-action models by generating an initial plan, then reusing and correcting its unexecuted suffix with one network evaluation per replan. Native action attention selects the execution horizon.

Critical strengths
  • A useful contribution to inference design: Treating predicted actions as persistent state provides a concrete way to exploit temporal continuity in generative control, without retraining or additional learned modules.
  • Strong practical payoff: Planner calls are 3.19× faster and episode-level planning is 2.47× faster, while RoboTwin success remains close to FastWAM: 91.27% versus 91.83%.
  • Evidence supports the central insight: Structural ablations support the importance of both temporal alignment and observation-conditioned correction. Evaluation across RoboTwin and LIBERO strengthens the empirical case.
Critical weaknesses
  • Generality remains unestablished: Experiments use one FastWAM backbone and simulated manipulation. Whether the inference principle transfers reliably across generative policies, embodiments, and physical environments remains open.
  • The boundary of useful reuse is unresolved: Abrupt changes or accumulated errors can invalidate the inherited plan. The framework lacks a mechanism for deciding when to abandon reuse and regenerate fully, limiting its applicability to less predictable environments (I guess the current method's succes can be dependent on the specific benchmark environments).
Questions for authors

Can unreliable inherited plans be detected online to trigger full decoding while retaining most of the speedup?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.