acceptodds
Under review as a conference paper at ICLR 2027

Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning

Abstract

Long-horizon world-model planning typically relies on autoregressive rollouts, where predicted states are repeatedly fed back into the model. This preserves temporal structure but creates a horizon-length sequential path and exposes later predictions to recursive decoded-state feedback. We introduce Parallel Predictive World Models (PPWM), which predict a finite-horizon trajectory in parallel while retaining causal interaction among future representations. Each horizon is conditioned on its causal action prefix, and future representations interact before decoding, separating temporal causality from state-by-state output recursion. We formalize this distinction by viewing autoregressive rollout as a causal trajectory map and identifying the decoded-state feedback pathway removed by PPWM. Across four visual-control tasks, PPWM achieves the lowest long-horizon prediction error and the highest Cross-Entropy Method (CEM) simulator success among the evaluated predictive interfaces. Meanwhile, PPWM achieves more than a average CEM planning speedup over the autoregressive LeWM baseline. These results suggest that accurate and efficient long-horizon world-model planning does not require state-by-state autoregression, but can instead be achieved through parallel causal trajectory prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
World Models, Long-Horizon Planning, Parallel Prediction, Causal Trajectory Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jugrOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Parallel Predictive World Models (PPWM) and a complementary trajectory refinement method. Compared with Fast-LeWM, PPWM introduces causal interaction among future-state representations before decoding. It also employs a scheduled-sampling-style training method called Multi-Anchor Self-Conditioned Rollout (MAR), which trains the model to continue predicting from its own predicted histories, reducing the training–inference mismatch during long-horizon prediction. For planning, it refines CEM-generated plans by jointly optimizing the action sequence and an auxiliary latent trajectory.

Critical strengths

The paper investigates a worthwhile research direction: how to construct world models that are both low-latency and sufficiently accurate, which is particularly important for their future use in planning.

Critical weaknesses

1. Limited architectural novelty and insufficient evidence for the proposed causal interaction. The main architectural contribution appears to be the addition of causal interaction among future-state representations to Fast-LeWM’s parallel prediction framework. However, causal attention among future representations is not, by itself, a clearly novel mechanism, and the paper does not sufficiently motivate why this additional interaction should improve upon action-prefix conditioning. More importantly, its contribution is not isolated through a direct, controlled ablation—for example, disabling cross-horizon interaction while keeping the remaining architecture, training objectives, and budget unchanged. Table 6 further complicates the interpretation: without MAR, PP-LeWM performs worse than Fast-LeWM at the 20- and 30-step horizons on Push-T. Although this does not establish that causal interaction is ineffective, it leaves unclear whether the reported gains primarily arise from the proposed architecture, the scheduled-sampling-style MAR training, or their combination.

2. Limited evaluation of out-of-distribution generalization. World models can perform well on familiar trajectories yet fail under distribution shifts, limiting their usefulness for planning, where candidate action sequences may depart from the training distribution. The paper does not explicitly evaluate such robustness—for example, training on trajectories collected by one behavior policy and testing on trajectories generated by another. Evaluations under shifted action distributions or unfamiliar initial states would help establish whether the reported gains extend beyond the training distribution.

Questions for authors

Please see the weaknesses above. In particular:

  1. Can you isolate the contribution of causal interaction among future-state representations through a controlled ablation, with and without MAR? How should we interpret PP-LeWM without MAR underperforming Fast-LeWM at longer horizons in Table 6?
  2. How were the training trajectories collected, and do the evaluations include shifts in behavior policies or action distributions relative to training?
Reviewer_dSxJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PPWM predicts a finite trajectory span in parallel. Horizon-specific queries receive causal action-prefix features and interact through a causal Transformer before decoding. Multi-anchor self-conditioned rollout training supports composition beyond the native span. The paper evaluates prediction and planning in frozen LeWM representations, adds CEM-initialized trajectory refinement, and tests prediction transfer to DINO-WM.

Critical strengths

The complete system has a substantial empirical case despite the simplicity of the architectural change. PP-LeWM has the lowest prediction MSE in every task/horizon entry of Table 1. Its CEM gains extend to simulator execution: Reacher terminal success rises from 64.33% to 96.67%, while average planning latency falls from 1235.2 to 374.5 ms on the reported hardware (Tables 2–3). Appendix D.3 also supplies an important control: PP-LeWM still leads on all four tasks when every model uses ordinary terminal-goal scoring. Thus, the planning result is not solely an effect of its tail-weighted objective.

The distinction between temporal causality and decoded-state recursion is presented precisely. Section 5 limits its analysis to the feedback pathway actually removed, rather than claiming to eliminate every source of compounding error. MAR and span-length experiments investigate long-horizon composition, and the DINO-WM result supplies supporting evidence beyond the main backbone, albeit on one prediction task.

Critical weaknesses

Section 6.6 overstates how cleanly the component analysis isolates future-query interaction. Fast-LeWM provides a useful parallel predictor reference, but it is not PP-LeWM with only that interaction removed. PP-LeWM also has a history-conditioned prefix branch and a fine-tuning curriculum (Appendix A.1), and the corresponding Fast-LeWM training and capacity controls are not explicit. Table 6 directly tests MAR, but not interaction within the otherwise fixed PP model. Without MAR, PP-LeWM's 30-step Push-T error is 0.239441, versus 0.208614 for Fast-LeWM. The results support the full recipe more strongly than they establish which architectural change accounts for its advantage.

The smaller planning and refinement improvements have uncertain reliability. The protocol uses 100 episodes and three planner seeds, but reports aggregate percentages without paired uncertainty or a per-seed breakdown. For example, common-goal Push-T terminal success changes from 27.67% to 29.33%, and Two-Room refinement improves terminal success from 55.00% to 57.67%. These differences should carry less weight than the large Reacher gain until their variation is shown. Refinement also does not uniformly beat action-only Adam: its Push-T terminal success is slightly lower.

Questions for authors
  1. With the PP-LeWM history/prefix modules, MAR, initialization, and training budget fixed, what changes when future queries cannot attend to other future queries? Please also specify Fast-LeWM's training budget and trainable capacity. This would resolve the main attribution concern.
  2. Can you provide paired episode-level or per-seed uncertainty for common-goal CEM and refinement, distinguishing planner randomness from independently trained-model variation?
Reviewer_uvcbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PPWM, which predict multiple future latent states in one forward pass. Future representations interact through causal attention, while decoded predictions are not fed back within a prediction span. The method also uses MAR for longer horizons and trajectory refinement for planning.

Critical strengths

The idea is simple and makes sense: predict the whole span in one pass but keep causal attention between future steps. The speedup over LeWM is real.

All prediction baselines share the same frozen LeWM latent space, so Table 1 is a fair comparison. I also liked that planning is evaluated by actually running the simulator.

The DINO-WM results also provide some evidence that the approach extends beyond the LeWM backbone

Critical weaknesses
  1. My main concern is that I can’t tell how much of the improvement comes from causal interaction among future queries, which is the main difference from Fast-LeWM. In Table 6, PP-LeWM w/o MAR is only about 3% better than Fast-LeWM at 10 steps on Push-T (0.0510 vs. 0.0526), and is worse at 20 and 30 steps. This makes me wonder whether MAR accounts for much of the improvement. I would like to see the same model with and without cross-horizon attention, keeping MAR and everything else fixed. Adding comparable self-conditioned training to Fast-LeWM would also help. Appendix D.2 mentions three design choices, but only shows the prediction-span sweep; there is no direct attention ablation.

  2. It is unclear whether the baselines received comparable training. PP-LeWM adds a history encoder and a prefix encoder, fine-tunes the dynamics backbone, and uses MAR. Did LeWM and Fast-LeWM receive similar training budgets and tuning? The Reacher results particularly need clarification. PP-LeWM has roughly 19 times lower MSE than LeWM over the first five steps, while Fast-LeWM is worse than LeWM. The gap is striking, and matched training settings would help establish whether it comes from the architecture.

  3. The main CEM results use different scoring functions, making it hard to separate the model improvement from the scoring change. PP-LeWM uses tail weighting with a task-specific coefficient. The goal-only comparison in Table 9 helps, but on Cube the advantage over Fast-LeWM shrinks to about 2.3 percentage points for both anytime and terminal success, with no uncertainty estimates. The speed comparison also needs the same scoring function: the paper states that Fast-LeWM’s dynamics are actually faster, but its self-consistency score requires extra model calls. Could the authors report CEM latency for all methods under goal-only scoring?

  4. The planning experiments never go beyond one prediction span. Open-loop planning uses H=K=10, and the closed-loop horizon decreases from 10 to 1. The paper tests prediction error beyond K, but not planning performance. I would like to see planning results at H>K, where the model has to reuse its own predictions as context for the next span.

Questions for authors

Please address the points raised in the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.