acceptodds
Under review as a conference paper at ICLR 2027

Go Further, Be Lighter: Recurrent Predictive Futures for World-Action Models

Abstract

World-action models (WAMs) repeatedly compute predictive futures for control, incurring substantial inference cost. Asynchronous execution reduces waiting, but the overhead of temporal correction can limit effective action-generation throughput. Crucially, the predictive state conditioning the executing action already anticipates the outcome of that chunk before the robot reaches the next control boundary. Our *key insight* is that even imperfect predictions capture chunk-end changes in task-relevant regions. This creates an opportunity to advance the predictive future into a further future for subsequent control. To this end, we introduce the **Further-Future Propagator (FFP)**, which treats the predictive state as a recurrent state and learns a future-to-further-future transition that advances the predictive state across control rounds guided by factual evidence. Action relevance shapes how information important for action generation is preserved and updated, while a recurrent manipulation context modulates how the predictive state evolves across control rounds. The composed state conditions the frozen native action pathway and becomes the next recurrent input. Before each transition, a Confidence module estimates the marginal value of obtaining a fresh native prediction from the WAM, enabling adaptive refresh instead of a fixed refresh schedule. Across three architecturally distinct WAMs on CALVIN and RoboTwin 2.0, FFP improves control–compute trade-offs. On Faster-WAM/RoboTwin 2.0 Easy, FFP achieves the effective action-generation throughput of Real-Time Chunking (RTC) while improving success rate from 76.1% to 92.2%. Across four real-world tasks with two WAMs, task-averaged completion time decreases by 34.3–39.3% relative to asynchronous baselines, with mean success rate maintained or improved.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world-action models, vision-language-action models, robot learning, predictive-state propagation, efficient inference, world models, asynchronous robot control

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SapsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the inference cost of world-action models, which condition action generation on a predicted future and must therefore recompute that prediction every control round, so that synchronous execution stalls the robot while asynchronous execution pays correction overhead and discards action steps whose scheduled times have already elapsed. Its premise is that the predictive state used to generate the currently executing chunk already describes the scene expected at the end of that chunk, so rather than being discarded it can be advanced into a further future. It proposes the Further-Future Propagator, which treats the model's action-facing predictive state as a recurrent state and learns a transition that carries it across control boundaries, grounded in compressed current and historical visual observations together with the committed action chunk, with an action-relevance attention bias over predictive positions and a continuous manipulation-regime state modulating feature evolution. A separate confidence module estimates the marginal benefit of obtaining a fresh native prediction, so refresh happens on demand rather than on a fixed schedule, and the native model stays frozen throughout. It is tested on three architecturally distinct world-action models in simulation on CALVIN and RoboTwin, and on a real robot across four manipulation tasks with two of those models, against matched asynchronous execution baselines, with component ablations, a static-reuse control, a no-refresh control, fixed periodic refresh schedules, and detailed per-module runtime accounting.

Critical strengths
  • Demonstrating the same mechanism on three architecturally distinct world-action models, whose action-facing predictive states are a hidden state, a future-video latent, and a set of fused key/value groups respectively, is stronger generality evidence than efficiency papers usually provide. Combined with a real-robot evaluation across four tasks and both carriers, and with per-module latency measurements reported as explicit arithmetic rather than claimed ratios, the empirical scope makes the core mechanism credible rather than tuned to one system.

  • The two negative controls do most of the conceptual work and are well designed. Reusing the predictive state unchanged for additional rounds nearly eliminates task performance, while running the learned transition with no native refresh also fails, which together isolate the claim that the predictive state must be actively advanced and periodically re-grounded rather than simply retained. Without these, the method would be indistinguishable from cache reuse.

Critical weaknesses
  • The headline comparisons are made against the asynchronous baselines rather than the synchronous reference the paper itself includes, which changes what the numbers mean. On the carrier with the largest reported speedup, the asynchronous baseline is far below synchronous execution on both success and usable action throughput, so the claimed success improvement is really a recovery of quality that baseline forfeits, and the method ends up roughly level with synchronous execution rather than ahead of it. The real-world results make this concrete, since average success falls on one carrier and stays below the synchronous reference on the other, which the abstract's claim about success being maintained or improved does not reflect.

  • No main-table success rate carries an error bar, and several load-bearing comparisons rest on differences of about one point, while an appendix depth sweep moves one carrier's success by ten points from a single configuration change. Combined with the fact that the always-refresh variant matches or beats the adaptive variant on task performance, and that the added module is very large for the carrier where it yields almost no quality gain, I cannot tell from the reported evidence which of the quality differences are real or whether the adaptive routing contributes anything beyond saving compute.

Questions for authors

-Against the synchronous reference rather than the asynchronous baselines, what does FFP actually buy on each carrier, and with what variability? Specifically, could you report success with confidence intervals over repeated runs for Native Sync, Always-Full FFP, and Adaptive FFP on the same tasks, so the roughly one-point differences in Table 1 and the four-point real-world drop on LingBot-VA can be distinguished from run-to-run noise?

Reviewer_Dg3ROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Further-Future Propagator (FFP), which treats the predictive state as a recurrent state and advances predictive representations across control rounds without running a full world-model prediction every time. It can saves the time for WAMs repeatedly running and also avoids the overhead of temporal correction. The experiments cover VPP on CALVIN, LingBot-VA and Faster-WAM on RoboTwin 2.0, and two WAMs on four physical manipulation tasks.

Critical strengths
  1. The paper proposes a very useful pipeline and has fundamental novelty in the proposed methods.
  2. The experiments are conducted over a range of different tasks, and including real-world experiment.
  3. The ablation study supports several design choices.
Critical weaknesses
  1. I think the paper may need to be more explicit about the improvement, e.g. how much of the reported throughput improvement arises from reduced computation, and how much from the effective-action counting convention and different asynchronous scheduling mechanisms?
  2. The time report for realworld just focuses on successful trials, but how much would be on average for every?
Questions for authors

Please address the questions in weakness.

Reviewer_TYe4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces the Further-Future Propagator (FFP) to reduce the inference cost of World-Action Models (WAMs) during asynchronous execution. FFP advances the action-facing predictive state across control rounds using a lightweight residual update conditioned on current visual evidence, executing actions, and manipulation stage. A Confidence module dynamically routes between FFP propagation and a full native WAM refresh. FFP is evaluated on VPP, LingBot-VA, and Faster-WAM across CALVIN, RoboTwin 2.0, and physical robots.

Critical strengths

The reviewer found the following strengths:

  • Regarding the method's novelties, the paper frames predictive futures as recurrent, updatable states, providing a practical, utility-guided alternative to rigid heuristic refresh schedules or purely stateless asynchronous wrappers.

  • The paper also validates the approach on three architecturally distinct models (hidden states, future-video latents, and fused K/V pairs), demonstrating broad conceptual applicability. The real-world deployment across four PiPER-X tasks successfully translates simulated efficiency into a 34.3–39.3% reduction in physical task completion time.

-Regarding presentation, the visual diagnostics motivating cross-boundary continuity are intuitive, and the evidence-routing-propagation pipeline is clearly formalized. Overall, it's enjoyable to read the paper.

Critical weaknesses

In my opinion, the paper still has the following main issues need to be addressed:

-The auxiliary "Stage" modulation requires a deterministic kinematic labeler that relies on hardcoded distance thresholds, motion windows, and gripper states. This heuristic assumes clean proprioception and well-defined interaction regions, limiting generalizability to unstructured environments. Furthermore, integrating FFP requires highly bespoke, carrier-specific coordinate mappings (e.g., routing eight fused K/V groups for Faster-WAM) rather than acting as a universal plug-and-play module.

  • (Missing Baselines) The empirical baselines are restricted to asynchronous wrappers like RTC and FDM Async. The evaluation omits comparisons against feature caching methods cited in the related work (e.g., VLA-Cache, DeepCache) or temporal ensembling techniques, which represent standard alternative paradigms for accelerating WAM inference.

-FFP uses the cached attention map from the executing chunk as a spatial prior for the next chunk. This backward-looking prior risks staleness during rapid phase transitions (e.g., instantly switching from reaching to transporting). Additionally, there is no evaluation verifying that the Confidence module can reliably trigger a native refresh to avoid catastrophic physical drift under out-of-distribution (OOD) visual perturbations. Indeed, this is not really a weakness, but rather a further experiment to validate the limit of the method. In summary, there is a risk of mechanism vulnerabilities that need to be examined carefully.

Questions for authors

I have the following questions:

  • How resilient is the Confidence module to OOD observations or sudden environmental perturbations? Does the system ever experience "catastrophic drift" where it repeatedly fails to trigger a native refresh despite large physical errors?

  • Did you observe performance bottlenecks at task phase boundaries (e.g., approach to grasp) where the relevant spatial regions suddenly shift, rendering the action-relevance prior stale?

  • How does FFP's control-compute trade-off compare to running the native WAM with DeepCache or VLA-Cache, or utilizing a temporal ensembling approach?

  • Could a self-supervised or vision-based approach replace the heuristic kinematic Stage labeler to improve generalization to in-the-wild datasets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.