acceptodds
Under review as a conference paper at ICLR 2027

-WM: Predicting Steps at Once for Long-Horizon Planning with World Models

31 views

Abstract

Latent world models have become an appealing approach to visual control, enabling goal-conditioned optimization at test time and generalization across tasks with a single model. In practice, however, planning success remains limited to short horizons, and gradient-based planners driven by autoregressive rollouts still underperform sampling-based methods, despite their potential to optimize action sequences efficiently through model gradients. We hypothesize that two properties of next-step prediction contribute to these limitations: autoregressive rollouts compound prediction errors, and planning gradients must traverse a chain of nested model calls. We study a simple alternative: predicting the entire future sequence in a single forward pass. We introduce -WM, a minimal extension to transformer-based JEPA world models: a multi-horizon training objective and input layout that let the predictor map a variable-length action sequence directly to the corresponding sequence of future latent states in a single forward pass. We apply -WM to three classes of JEPA world models and evaluate on four environments. Overall, -WM improves gradient-based planning by 37% and sampling-based planning by 15% over the existing next-step world models it extends. Our analysis shows that the gains cannot be explained by parallel rollout alone: improvements persist under autoregressive inference and coincide with substantial changes in the learned representation. Parallel prediction additionally reduces long-horizon inference cost. Together, -WM offers a simple, broadly applicable path to more accurate and efficient long-horizon control. Videos are available at https://anon-k-wm.github.io.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, JEPA, Planning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_6LLHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes k-WM, which lets JEPA world models to predict the k-step future in a single step to remove autoregressive generation and to deal with compounding error, improving gradient based method for goal conditioned planning. Applied on DINO-WM, finetuned DINO and LeWM, k-WM presents improvements of CEM and GD on open-loop planning performance and efficiency. They also present representational analysis and rollout visualizations with less compounding errors.

Critical strengths
  • Clear motivation and methodology. JEPA latent world models and planning suffered from representation collapse and compounding error during rollout making it hard for gradient based optimization to work properly. This work proposes minimal changes to the world model to support multiple horizon predictions and variable length inference, significantly improving gradient based planning performance in challenging scenarios.
Critical weaknesses
  • Distinction relative to Fast-LeWM. The core mechanism, prediction of all future latents from an action sequence over multiple horizon, is already in Fast-LeWM, claiming similar long-horizon improvements, faster planning, and representational gains. The main distinction is how they implemented such design and their additional analysis. More matched comparison on the design choices seems to be required to explicitly compare the two methods.
Questions for authors
  • Encoder-matched comparison. Can you also present results using the prior baseline methods (LeWM and Fast LeWM) with a DINO or ViT-S encoder to show a comparable architecture with k-WM with DINO-WM/DINO-FT? While planning performance is one of k-WM’s clear distinctions, it seems Fast-LeWM already matches performance of k-WM LeWM concat with a comparable backbone scale.
Reviewer_YDMhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies long-horizon planning with latent JEPA-style world models. The authors argue that conventional next-step world models suffer from two issues during planning: autoregressive rollout compounds prediction errors, and gradient-based planners must backpropagate through a long chain of model calls. They propose k-WM, a simple multi-horizon extension that predicts multiple future latent states from the current latent context and an action sequence in a single forward pass.

The method is evaluated on three world-model variants (DINO-WM, fine-tuned DINO, and LeWM), four control environments (Push-T, Maze, Cube, and Reacher), and both CEM and gradient-based planning. The paper reports improved planning performance, particularly for end-to-end trained encoders and gradient-based planning, together with lower long-horizon inference cost. It further separates training horizon from inference chunk length and analyzes how multi-horizon supervision changes the learned latent representation.

Critical strengths
  1. Simple and broadly applicable formulation with a meaningful empirical effect. k-WM is a minimal modification to existing transformer-based JEPA world models, yet it consistently improves several planning settings, especially gradient-based planning with end-to-end trained encoders. The simplicity of the formulation makes the result practically relevant and easy to transfer to related latent world-model architectures.

  2. Strong experimental analysis beyond final task performance. The paper evaluates three model families, four environments, and two planners, and includes useful ablations over training horizon and inference chunk length. Importantly, it also studies rollout quality, planning speed, and representation geometry. The experiments separating multi-horizon training from parallel inference are especially valuable because they suggest that the gains are not solely an implementation-level speedup.

Critical weaknesses
  1. The main matched comparison does not fully isolate the effect of multi-horizon training. k-WM receives supervision over multiple future targets per training sequence, while the next-step baseline uses shorter-horizon supervision. It is therefore unclear whether the comparison is matched in terms of the number of supervised targets, observed frames, tokens, and training FLOPs. In addition, the long-horizon LeWM results compare k-WM against the original AdaLN LeWM rather than the otherwise more closely matched LeWM-concat next-step baseline. These factors weaken the claim that the reported improvements are specifically caused by the proposed multi-horizon objective.

  2. The evidence for the “long-horizon” and gradient-conditioning claims is not yet fully convincing. The main planning evaluation extends only to H=10H=10, which is close to the effective prediction range of the default multi-horizon model, so the paper does not strongly test repeated composition far beyond the training horizon. Moreover, although improved gradient conditioning is a central motivation, the paper does not directly measure gradient quality, Jacobian conditioning, or optimization behavior..

Questions for authors
  1. Can the authors provide a baseline that is matched to k-WM in terms of the number of observed frames, supervised future targets, training tokens, and preferably training FLOPs? This would help determine whether the gains come from multi-horizon structure rather than simply additional supervision.
  2. Can the authors report long-horizon results for the matched LeWM-concat, k=1k=1 baseline, and ideally evaluate horizons substantially beyond the training/effective one-pass horizon (e.g., H=20H=20 or larger)? This is important for assessing whether k-WM remains advantageous when multiple prediction chunks must be composed.
  3. Do the reported seeds correspond to independently trained world models or only planning/environment seeds? Please clarify the number of independent training runs and whether evaluation trajectories/goals are strictly disjoint from world-model training data.
  4. Since improved gradient conditioning is a central motivation, can the authors provide direct evidence such as gradient norms versus horizon, optimization convergence curves, Jacobian conditioning, or another diagnostic showing that k-WM actually improves the optimization landscape seen by the gradient-based planner?
1 comment

Thank you for the detailed response and additional experiments. While these additions address several of my concerns, I still do not believe that the paper's novelty and contributions meet the bar for ICLR, so I will maintain my original rating.

Reviewer_Ns75OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes k-WM, which uses multi-horizon supervision and a modified input layout to predict future latent states in parallel. It evaluates both gradient-based and sampling-based planning and studies the effects of training horizon and inference chunk length. I consider the improvement useful and welcome comparisons with LeWM and Fast-LeWM. However, changes to the training and evaluation configurations make the reported gains difficult to interpret.

Critical strengths
  1. The method is simple, reasonably motivated, and practically relevant. Improving multi-step latent prediction is a useful contribution to planning with world models.

  2. Evaluating both gradient-based and sampling-based planners, while separating the effects of training horizon and inference chunk length, provides useful insight into how the method behaves under different planning settings.

Critical weaknesses
  1. The experiments introduce too many changes beyond the proposed method. The main contributions concern the prediction formulation and training objective. However, the experimental setup also changes the training data, goal sampling, and planning protocol, removes model predictive control (MPC), and introduces proprioceptive inputs. The necessity of these changes and their connection to the core contributions are insufficiently explained.

  2. The comparison setup and presentation of results may cause confusion. Table 1 states that prior methods are re-evaluated under a unified planning protocol, but were LeWM (AdaLN), Fast-LeWM, and Straightening retrained for these comparisons? If not, the comparisons involve differences in training data and potentially in access to proprioceptive inputs. Moreover, some of the strongest results in Table 1 use autoregressive inference, whereas Figure 5 emphasizes the efficiency of parallel prediction. Although Figure 5 evaluates Push-T rather than Reacher, success rates and runtimes should be reported together under the same planning configuration, so readers do not interpret the best results from different configurations as simultaneous speed and performance advantages.

  3. Grid search over hyperparameters. For LeWM, tuning the SIGReg weight separately for each task using the appendix’s grid search is inappropriate; using a shared weight across tasks, as in standard LeWM [1], would be more convincing. Reporting results with a shared SIGReg weight and the same number of training epochs across tasks would help establish whether the gains consistently arise from the multi-horizon training objective.

  4. Baseline performance needs verification. In Table 1, LeWM (AdaLN) and Fast-LeWM achieve only 24.7% and 17.3% CEM success rates on Reacher, respectively—close to random-policy performance and far below the results reported over more random seeds in prior work [2, 3]. Fast-LeWM also shows a substantial performance drop on Cube. I understand that the authors use a single planning call followed by open-loop execution, but discrepancies of this magnitude still warrant investigation and explanation.

  5. The role of parallel inference in gradient-based planning remains insufficiently explained. A key motivation is to avoid gradient degradation through long autoregressive computation chains, yet direct evidence on gradient quality or optimization behavior, such as action-gradient stability and planning-loss convergence curves, is missing. In Table 2, Reacher’s gradient-descent (GD) planning success rate drops from 68.7% at (d=1) to 12.7% and 22.0% at (d=2) and (d=4), before recovering to 60.7% at (d=6). This indicates that reducing recursion depth alone does not guarantee easier optimization. The strong DINO FT results on Reacher in Table 1, as well as the Reacher results in Figure 3, also use autoregressive inference. These results support the value of multi-step training but do not directly establish the advantages of parallel inference.

References:

[1] Maes et al. LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. 2026. https://arxiv.org/abs/2603.19312

[2] Gao and Xu. Fast LeWorldModel. 2026. Appendix A, Table 5. https://arxiv.org/abs/2606.26217v2

[3] Zhao et al. Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models. 2026. Table 1. https://arxiv.org/abs/2605.09241

Questions for authors
  1. Given that the proposed planning protocol brings baseline performance close to that of a random policy, could the authors conduct a paired evaluation using LeWM’s training data and planning protocol, keeping AdaLN action conditioning, input information, and other training settings unchanged, and introducing only k-WM’s multi-horizon extension and dense multi-horizon loss? This would help quantify the performance gains that k-WM’s extension provides over standard LeWM (AdaLN).

  2. Why does Reacher’s GD planning performance vary so strongly and non-monotonically with rollout chunk length in Table 2? Do failures occur because the planning loss does not decrease sufficiently, or because the resulting actions fail to complete the task despite a reduced loss?

  3. Please verify the model checkpoints, simulator-version requirements, and any relevant fixes in the latest LeWM and Fast-LeWM repositories to ensure that the correct, up-to-date baselines are used.

  4. Please specify the inference mode used for each result in Table 1 and separately report success rates and runtimes for autoregressive rollouts and parallel multi-step prediction. Parallel multi-step comparisons should use the same rollout chunk length. Overall, Table 1 is highly misleading: it compares results involving different training data, planning configurations, and input information, making it easy to combine the best performance and speed advantages from different settings.

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

The paper proposes a tweak to transformer-based Jepa-style models which allows them to predict k-steps into the future immediately, rather than having to unroll a 1-step model k times. The approach is evaluated as a wrapper around 2 Dino and one LeWM architectures, in 4 relatively small-scale tasks (the most complex being Reacher). The results are as expected (k-step prediction can avoid compounding errors of 1-step rollouts, thereby helping planning in most cases).

Critical strengths

WMs are an important topic and this paper is playing around with an idea that, while explored before, has not been studied enough in transformer architectures. Experiments are executed competently.

Critical weaknesses

There is little new in this paper. Multi-time scale models in RL go back to Sutton's 1995 paper and subsequent work on option models by Sutton Precup et al. This is essentially the same idea, just done on transformers, and none of that work is cited. The motivation and experimental results back then were very similar to the ones presented here. Therefore, the novelty is mainly in doing this with Jepa-style architectures, which seems like a minor variation. This would be fine if the results were really strong, but they are actually middling, with the results still degrading significantly with longer time horizons (see eg Fig 3). The domains used for the evaluation are relatively toy, and open-loop rollouts are only good for deterministic domains (therefore applicability is limited).

Questions for authors

Please try a larger-scale experiment on a domain that actually needs longterm planning (eg Go, Crafter), in order to make a stronger case for the proposed approach. Fig 4 - what about error bars? Please add learning curves for the model and/or a sense of data costs associated with the proposed method (does it need more data to train)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.