acceptodds
Under review as a conference paper at ICLR 2027

Investigating Priors for Informed Latent Planning in World Models

Abstract

World models leveraging the Joint Embedding Predictive Architecture (JEPA) are trained to reach goals zero-shot through latent planning. Most work starts from an uninformed Gaussian prior for planning, while a few recent works show the advantage of learned priors. However, the effectiveness of different priors is not thoroughly studied. In this work, we investigate three priors learned by different policy heads, i.e., a deterministic head (Behavior Cloning), a Gaussian one, and a generative one (Flow Matching). We evaluate these priors on success rate together with planning latency across four world model backbones and six robot environments. Our results show that learned priors can improve the success rate by a large margin, e.g., by percentage points in the Cube environment. Moreover, learned priors exhibit computational efficiency, reducing planning time per action sequence by a factor of to across all environments, highlighting their practical utility. We show that while no single prior dominates in the short-horizon setting, Flow Matching learns a more robust prior most capable of challenging long-horizon planning tasks. Based on our findings, we distill guidance on choosing between priors, and other algorithm settings such as replanning cadence, to inform the design of future latent planning systems for real-world applications.

Primary area:
foundation or frontier models, including LLMs
Keywords:
World Models, Latent Planning, Robot control, Generative Modeling, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zYY8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates learned priors for latent planning in JEPA-based world models. The authors systematically compare three types of priors—deterministic Behavior Cloning, Gaussian, and generative Flow Matching—across four world-model backbones (LeWorldModel, PLDM, DINO-WM, and its Straightened variant) and six control environments (Two-Room, Push-T, OGBench-Cube, Wall, PointMaze-UMaze/Medium). Experiments show that learned priors significantly improve success rate and reduce planning latency when trained on goal-directed demonstrations, with Flow Matching proving most robust under frequent replanning and long-horizon settings. The paper also analyzes how sample size, replanning cadence, and goal-distance weighting affect prior effectiveness.

Critical strengths
  1. The trade-off between learned priors and uninformed search in latent planning is a central yet under-explored question. This work provides the systematic comparison of three priors with varying expressiveness (BC, Gaussian, Flow Matching) within a unified framework across multiple world models and planning settings.

  2. By having all three priors share the same FiLM-conditioned residual MLP backbone and differ only in their output distributional form, the authors effectively isolate the variable of interest. The uniform sample budget and shared planning objective further ensure a fair comparison.

Critical weaknesses
  1. Llimited experimental coverage undermines generalizability. All six environments are continuous-control tasks (2D/3D navigation and manipulation) with relatively low-dimensional action spaces. This raises two major concerns:

    • Restricted environment scope: The paper completely omits discrete-action environments (e.g., Atari), high-dimensional visual tasks, or real-world robotic scenarios. The behavior of JEPA world models and the effectiveness of learned priors may differ drastically in discrete-action or sparse-reward settings, leaving the generality of the conclusions unclear.
    • Narrow algorithmic comparison: The baselines are limited to CEM and gradient descent. Notably, the authors do not compare against—or even discuss—MCTS-based planners such as MuZero or UniZero, which also rely on learned priors (policy and value networks) to guide planning and have demonstrated strong performance on complex decision-making tasks. Given the authors' claim of being the first to thoroughly study the advantage of different priors, the absence of any comparison or discussion with these methods substantially weakens the contribution.
  2. Limited discussion of prior acquisition sources. All priors in this work are trained from the same offline demonstrations via behavior cloning or generative modeling. However, prior acquisition is not limited to demonstration data. For instance, PriorZero and related works have explored using Large Language Models as external knowledge sources to generate task priors or subgoal decompositions, offering structured guidance without task-specific demonstrations. The authors do not discuss how such "non-data-driven priors" relate to, complement, or compete with the learned priors studied here. Given the rapid growth of LLM-for-robotics research, this omission makes the discussion of prior design feel incomplete.

Questions for authors
  1. Have the proposed priors been validated in discrete-action environments (e.g., Atari)? If not, do the authors believe that continuous-space generative priors like Flow Matching can naturally generalize to discrete-action settings, or would fundamental architectural changes be required?

  2. Compared with MCTS-based methods such as MuZero / UniZero, how do the sample efficiency and asymptotic performance of the proposed CEM-based informed planning compare? Under what circumstances should one prefer sampling-based priors over tree-search-guided priors?

  3. Given that demonstration quality critically determines prior effectiveness, have the authors considered leveraging LLMs or other external knowledge sources to construct or augment priors, thereby reducing reliance on high-quality demonstrations?

  4. Reasonableness of the long-horizon extending baseline. In the long-horizon experiment (Figure 2), short-horizon priors (H=5H=5) are extended to H=10H=10 by padding with Gaussian-sampled actions. This baseline seems artificially weak—why not extend the chunk autoregressively using the prior itself (e.g., feeding its own output back as the next conditioning input)? If a stronger extending strategy were used, would the relative advantage of retraining at H=10H=10 remain as pronounced?

3 comments

We thank the reviewer for their feedback and address the concerns below.

W1.1: Limited experimental coverage: No discrete-action, high-dimensional visual tasks, or real-world robotic scenarios.

Our study follows the experimental settings of current research on JEPA-based world models and latent planning (LeWorldModel, DINO-WM, temporal straightening): continuous control of differing complexity, from 2-D navigation and pushing to 5-D arm control in OGBench-Cube. Discrete action spaces such as Atari would require changing the priors' output parameterization and CEM's Gaussian refit (see Q1). They also motivate tree search rather than sampling-based planning (see W1.2). Regarding high-dimensional visual tasks, all six environments are visual control tasks. The world models plan from 224×224 RGB observations (about 150k input dimensions, about 5× the standard Atari input of 84×84×4 grayscale frames) in latent spaces of 192 (LeWM, PLDM) to 1,568 (DINO-WM) dimensions. The planners search over action chunks of 50–250 dimensions with only 300 samples. Real-world robotic scenarios remain a limitation, which we state in the manuscript.

W1.2, Q2: Absence of MCTS-based planners such as MuZero or UniZero.

MuZero and UniZero learn reward, value and policy from online interaction and use the policy as a search prior, typically over discrete actions. Our setting differs in each respect: The world models are trained on reward-free offline data, goals are given as images at test time (zero-shot), there is no value function, and actions are continuous chunks. Sampled MuZero extends MuZero to continuous actions and MuZero Unplugged to offline data, but both still require a reward signal to learn value and policy targets, which our reward-free, goal-image-conditioned setting does not provide. Without rewards, these methods do not apply out of the box, so sample efficiency and asymptotic performance are not directly comparable. Sampling-based priors are preferable for continuous, high-dimensional action chunks, reward-free goal-conditioned tasks, and frozen pre-trained world models. Tree-search-guided priors are preferable for discrete actions with a reward signal and online interaction. We will add a discussion of MCTS-based planners with learned priors, including MuZero, UniZero and these variants, to the related work.

W2: Limited discussion of prior acquisition sources.

We thank the reviewer and will add a discussion of non-data-driven priors. LLM priors such as PriorZero act on discrete, language-describable actions or subgoals, whereas our priors propose low-level continuous action chunks. The two are orthogonal: an LLM could propose subgoals that condition our action priors. Like previous work on JEPA-based world models, we limit the scope to image observations and continuous control. LLM priors are better suited to tasks whose actions or structure can be described in language, such as text games or high-level task decomposition.

Q1: Validation in discrete-action environments?

They have not been validated in discrete-action environments. The hindsight training and the conditioning on start and goal latents would carry over unchanged, but the output parameterization would have to change. The Gaussian and BC priors would need to predict categorical distributions over action chunks, and Flow Matching and CEM would need to be replaced by their discrete versions. Please see also W1.1.

Q3: LLMs or external knowledge to reduce reliance on demonstrations.

We deliberately used the world models' own training data, which comes at no extra cost and varies in quality. Even so, the priors reduce planning time and often raise success rates. Our data-efficiency study (App. Figure 9) further shows that training priors only on 25% of the demonstrations retains nearly the full benefit, which reduces the required size of demonstration data. Moreover, our new data-quality sweep on Two-Room (see our reply to Reviewer uP93, Q2) shows that priors trained on data that is only 25% goal-directed already reach near-ceiling success. We agree that LLMs or other external knowledge sources are a promising way to reduce this reliance further (see also W2).

Q4: Gaussian padding is an artificially weak extension baseline and extending the chunk autoregressively would be stronger.

We thank you for suggesting a stronger extension strategy. In short, autoregression improves on Gaussian padding in all 24 configurations and closes a substantial part of the gap, but retraining at H=10 stays ahead for Flow and for all priors combined with CEM, so the advantage of retraining persists, although it becomes smaller. The full results are in our general response.

Thank you for the additional experiments and clarifications, which address several of my concerns. Further experiments on non-data-driven priors would still help broaden the discussion of prior design, although I understand that they operate at a different level from the proposed low-level action priors. Likewise, although MuZero/UniZero cannot be applied directly to the current setting, preliminary results with an adapted MCTS baseline would further strengthen the paper’s empirical claims.

I still have a question regarding the paper’s core contribution. That a learned prior trained on goal-directed demonstrations substantially improves success rates is relatively intuitive. Although the general response already provides several useful practical guidelines, could the authors further clarify which findings go beyond this intuition, which lessons transfer robustly across environments and world models, and what their conditions and limitations are? This would help sharpen the paper’s reusable takeaways and distinct contribution.

We thank the reviewer for the follow-up and are glad that our reply addressed several concerns. As for the remaining question, several of our findings run counter to what intuition would suggest, four of which stand out. First, one would expect the most expressive prior to work best, yet the simplest, deterministic prior often matches or beats Flow. Flow’s advantage only emerges when the planner can choose among many candidates, with dense replanning, or at long horizons. Second, one would expect a prior to be useful only if it was trained on goal-directed data, yet priors trained on random-walk data did not mislead the search. Used to initialize CEM, they matched or improved on it. Third, one would expect the choice of prior to matter most, yet replanning cadence and cost weighting affect success far more. Fourth, one would expect the best planning objective to depend only on the task and world model, yet under frequent replanning our experiments suggest that it depends on how candidates are proposed: Flow tends to prefer weighting all steps equally, the unimodal priors favor early steps, and CEM favors later ones.

Among our findings, the following three proved most robust across world models. Seeding CEM with a Flow or Gaussian prior matches or improves on CEM, whereas the deterministic prior is less reliable. A prior alone reliably replaces search and cuts planning time several-fold when the demonstrations are goal-directed. Otherwise, it is best used to seed CEM. On Push-T, Flow is the most robust prior under dense replanning and at long horizons under all four world models, provided the sample budget is sufficient. With few candidates, the deterministic prior is preferable. These lessons are limited to simulated continuous-control tasks, and the priors are trained per world model and environment.

Our contribution therefore goes beyond showing that priors trained on goal-directed demonstrations help. It shows when priors help even without such demonstrations, which prior to choose, and how to configure the planner around it.

Reviewer_uP93OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work explores various priors for CEM and GD based planning in JEPA-based world models. The authors motivate the use of priors and describe the three respective prior-proposing methods well. There are three priors:

  • Behavior Cloning, where a pre-trained, goal conditioned policy proposes a single action chunk and this chunk is used for initializing planning rollouts. To maintain diversity of actions a guassian noise is added to perturb these chunks.
  • Gaussian head, where the model predicts parameters of a normal distribution, from which the action chunks are sampled.
  • Flow Matching which gradually transforms a random Gaussian noise into action chunks. This has an advantage that FM can represent multimodal distribution and therefore is the most expressive.

The authors then ask multiple questions:

  1. Do learned priors beat uninformed planners in success rate and planning time?
  2. Does combining prior + planning help compared to using either alone?
  3. Do priors trained on short horizons transfer to long horizons?
  4. Whats the relationship between planning time and success rate ?
  5. How do various hyperparameters affect performance?
Critical strengths
  • The authors choose reasonable setups with very similar parameter counts and sample budgets which tries to eliminate any bias towards any specific methods.
  • The set of questions that the authors pose have practical value.
Critical weaknesses

The main weakness with this paper is that it is hard to take away a reasonable lesson from the conducted experiments.

Table 1 evaluates LWM and PLDM on different sets of environments than DINO-WM. For LWM and PLDM it seems like priors help, while on DINO-WM they do not. But it is not clear whether the performance difference is because of the models, priors, or different environments. Authors offer an explanation but it cannot be verified, as the WMs, environments and data quality are fully confounded.

For table 2, most rows are within noise, which I think could be solved by more seeds as they seem to be cheap. Also the claim about "sota" numbers is overstated (only relative to their own uniformed baselines, not other work) and the gains of FM are highlighted compared to FM alone, not compared to CEM, which should be the main comparison, as the paper is about evaluating priors for methods like CEM.

The figure showing that retraining helps feels obvious and the baseline is very weak (just gaussian padding).

The replanning cadence and cost weighting results are simply not surprising. R=HR=H with b=∞b = \infty is exactly the task that is evaluated, R=HR=H with finite bb is just a mismatched objective. Same for R<HR < H, since the first action is executed assuming that another H−1H-1 actions will happen after, so the policy overshoots.

The authors mention that most experiment take under a minute to at most 11 minutes, yet they only report averages over only 3 seeds. The results would be clearer with more seeds, which seems relatively cheap to run.

Questions for authors
  1. What is the central takeaway? If a practitioner wants to take a lesson for his following work, what should it be?
  2. In Table 1, the world model, environment and data quality change together. Could you vary one at a time (e.g., LWM/PLDM on Wall and PointMazes) to show what actually determines when priors help?
  3. Authors put three methods on an expressivity ladder, but on the short horizons no prior is consistently better. The only advantage of Flow is on long-horizon Push-T. Are there more settings where the stronger expressivity/multimodality is necessary to show the need for more expressive priors?
2 comments

We thank the reviewer for their feedback and address the concerns below.

W2: Confounding of world models, environments and data quality.

We agree that our six environments are not shared by all world models. Push-T, however, is used with all four world models with the same success criterion and demonstration data, and it therefore isolates the effect of the world model. On Push-T, the priors behave the same way under every world model: Priors alone are roughly on par with CEM, Priors seeding CEM improve on CEM, and at the long horizon, Flow outperforms CEM under all four world models. Thus, the world model is not what decides whether a prior helps. We attribute the difference in Table 1 mainly to the environments’ data. Two-Room and Cube come with goal-directed demonstrations, whereas Wall and the PointMazes come with random-walk data. We isolate this effect in a controlled data-quality sweep on Two-Room, varying only the share of goal-directed demonstrations under a fixed world model and environment (see Q2). The sweep isolates a clear effect of data quality: goal-directedness is learned and helps, and 25% goal-directed data already suffices for near-ceiling success.

W3: The "claim about 'sota’ numbers is overstated” and "the gains of FM are highlighted compared to FM alone, not compared to CEM, which should be the main comparison”.

We agree with you and will polish the manuscript to accurately describe the benefit of the action priors and rewrite it around the Prior+CEM vs. CEM narrative (e.g. Flow+CEM achieves 12 to 14 pp higher success rates than CEM on Push-T using the DINO-WM variants).

W3, W6: Run experiments with additional seeds.

We reran the experiments behind Tables 1 and 2 with 7 seeds and report the results in our general response. None of the conclusions has changed.

W4: “The figure showing that retraining helps feels obvious and the baseline is very weak (just gaussian padding).”

Gaussian Padding is not the baseline. The baseline is CEM, which is not weak. With this figure, we want to highlight that despite the horizon mismatch, short-horizon priors with Gaussian padding reach accuracy comparable to CEM on Push-T. I.e. Gaussian padding offers a retraining-free option that stays on par with CEM, and clearly exceeds it once combined with CEM. Following Reviewer zYY8, we additionally evaluated a stronger, retraining-free extension that rolls the H=5 prior out autoregressively. Our general response contains the results. Extending the short-horizon priors autoregressively improves on Gaussian padding in all configurations. Autoregression bridges a substantial part of the gap between the Gaussian-padding short-horizon priors and the priors trained natively at H=10. Compared to CEM, the autoregressively extended priors alone outperform it in 11 of 12 configurations, and seeding CEM with them improves on CEM in all configurations, by 9.7 to 19.9 pp.

W5: “The replanning cadence and cost weighting results are simply not surprising.”

We agree that R=H with a terminal embedding cost (b=inf) performing best is expected. It serves as a sanity check. However, there are two non-obvious findings for R<H that we want to highlight:

  • (i) The best cost weighting b depends on the planner (CEM consistently prefers weighting later embeddings higher with b=2, Flow mostly prefers equal weightings with b=1 and BC consistently and Gaussian mostly prefer overweighting earlier embeddings with b=0.5)
  • (ii) the priors respond very differently. Under the best b, the Flow is very robust under frequent replanning with the success rate changing by -6 to +1 pp from R=5 to R=1, while the Gaussian prior drops by 8-26 pp and BC by 17-36 pp.

W1, Q1: Central takeaways for Practitioners.

Our central takeaway is that whether and how to use a learned prior depends more on the data and the planning setting than on the prior’s expressivity. Concretely, a practitioner should: (1) use a prior alone if the demonstrations are goal-directed; even the simplest deterministic prior then matches or beats CEM at a fraction of the planning time; (2) use the prior to seed CEM rather than replace it if the data is undirected; (3) choose Flow when replanning densely, planning over long horizons, or sampling many candidates, and a cheaper unimodal prior otherwise; (4) tune replanning cadence and cost weighting together with the planner and train the prior at the horizon it will be used at; and (5) compare planners across several sample sizes, as conclusions can reverse. We summarize these in the general response and will add them as a dedicated guideline paragraph.

Q2: “In Table 1, the world model, environment and data quality change together. Could you vary one at a time (e.g., LWM/PLDM on Wall and PointMazes) to show what actually determines when priors help?”

Push-T already varies the world model while keeping environment and data fixed (see W2). To vary the remaining factor, we kept world model and environment fixed and changed only the quality of the prior’s training data. The tables below show the success rate (averaged over seven seeds) on Two-Room with priors trained on different mixtures of the original demonstrations and length-matched, random-walk episodes. 100% indicates that the priors were only trained on original demonstrations while 0% indicates that the priors were only trained on random-walk episodes. We can see that even with pure random-walk episodes as training data, the priors are able to match or beat the much slower CEM algorithm. We attribute this to the similarity to the original demonstrations which are also noisy (but additionally goal-directed) trajectories. Despite this, we can see a pronounced gap between 25% and 0% highlighting that the goal-directedness of the original demonstrations is learned and helpful. The table also indicates that 25% goal-directed demonstrations suffice for near-ceiling success.

Impact of data quality on Two-Room with LWM (CEM: 83.93 ± 3.98)

Planner100%75%50%25%0%
Flow96.50 ± 1.4497.00 ± 1.5096.43 ± 1.5795.71 ± 1.5590.29 ± 1.98
Gaussian98.50 ± 0.7198.14 ± 1.1198.07 ± 1.1098.07 ± 1.0295.93 ± 1.74
BC99.86 ± 0.2499.93 ± 0.1999.79 ± 0.2799.86 ± 0.2496.29 ± 0.76
Flow+CEM100.00 ± 0.00100.00 ± 0.00100.00 ± 0.0099.93 ± 0.1999.43 ± 0.53
Gaussian+CEM100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.0099.14 ± 0.56
BC+CEM100.00 ± 0.00100.00 ± 0.00100.00 ± 0.0099.93 ± 0.1998.14 ± 1.11

Impact of data quality on Two-Room with PLDM (CEM: 95.29 ± 2.02)

Planner100%75%50%25%0%
Flow97.00 ± 1.2697.00 ± 0.9696.00 ± 1.7397.07 ± 0.9393.14 ± 2.10
Gaussian98.64 ± 0.6998.79 ± 0.5798.93 ± 0.7998.07 ± 1.1395.43 ± 1.40
BC99.79 ± 0.2799.93 ± 0.1999.93 ± 0.1999.86 ± 0.3896.21 ± 1.65
Flow+CEM100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.0099.86 ± 0.24
Gaussian+CEM100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.0099.86 ± 0.24
BC+CEM100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.0098.93 ± 0.84

Q3: “Are there more settings where the stronger expressivity/multimodality is necessary to show the need for more expressive priors?”

Yes. One example is under dense replanning: at R=1, Flow outperforms the unimodal priors by 16-40 pp and beats CEM under all four world models (Tables 3 and 13). A second example is larger sample sizes: at N=500, Flow outperforms both unimodal priors under all four world models (Table 12). Together with long-horizon Push-T, where Flow leads the unimodal priors by 7–14 pp under all four world models (Table 8), this gives three settings where expressivity pays off: dense replanning, large sample budgets and long horizons.

Reviewer_mg9YOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates a method to boost planning algorithms by introducing an action prior. The planning algorithm used across the latent world model literature is an instantiation of black-box optimization, and it starts from a randomly initialized action trajectory. What the authors focus on in this work is leveraging an action prior instead of random initialization, so that optimization can start from a region around feasible actions. While this idea is not new and has already been widely explored in previous work, the core distinction of this work is its investigation of the action prior itself, such as which formulation of the policy model works best. However, the resulting tendency is not very surprising: the policy model with the strongest capability (i.e., the flow-matching policy) is also the one that works best as an action prior.

Critical strengths

The paper is easy to follow. The writing is clear overall, and I had no trouble following the authors’ claims.

Critical weaknesses
  • The conclusion is trivial: the paper concludes that a flow-matching-based action prior works better than deterministic or Gaussian-based ones. This result is easily predictable, since the architecture that works best for policy learning also wins in the planning setup. I don’t think the current experiments and conclusion deliver a meaningful technical contribution given the bar of ICLR.
  • Experimental setup: The experiments are conducted in simple, synthetic environments that do not possess properties such as contact richness or high-dimensional action spaces, where CEM struggles (the current environments are ones where CEM already does fairly well). Recently, there has been work on planning with action priors in more challenging simulation environments or real-world setups [1,2,3]. I recommend the authors check more recent work and provide more in-depth insight into the role of action priors in planning. The current conclusion is somewhat surface-level, as lessons from BC can be directly projected onto planning.

[1] Jain, Arnav Kumar, et al. "A smooth sea never made a skilled sailor: Robust imitation via learning to search." Advances in Neural Information Processing Systems 38 (2026): 62676-62707.

[2] Seo, Gawon, Dongwon Kim, and Suha Kwak. "ACID: Action Consistency via Inverse Dynamics for Planning with World Models." arXiv preprint arXiv:2607.02403 (2026).

[3] Chun, Junha, Youngjoon Jeong, and Taesup Kim. "Sparse imagination for efficient visual world model planning." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors
  • Naming the deterministic prior “BC” is misleading, since every action prior listed here is technically an instantiation of BC.
  • I think a more realistic setup would be to use a much more powerful action prior, instead of training it jointly with the world model via hindsight relabeling. For example, one could use a larger-scale VLA and use the world model as a verifier for inference-time search ([2, 3] is already doing this).
1 comment

We thank the reviewer for their feedback and address the concerns below.

W1: The conclusion that the policy model with the strongest capability works best as a prior is not very surprising.

Respectfully, our main finding is not that Flow generally works best. Our results show that the priors separate far less than their expressivity suggests, and none dominates in short-horizon planning. Two examples:

  • Goal-directed demonstrations (Two-Room, OGBench-Cube): the simplest prior, BC, reaches 99.9–100% on Two-Room and 99.4–99.5% on Cube under both LWM and PLDM. That matches or exceeds Flow (96.0–99.1%) at about half its planning time (Table 10). With PLDM, this also holds at the long horizon on Cube (Table 8).

  • Sample size (Push-T, Table 12): Flow's advantage depends on how many candidates the world model can choose from. At N=30, Flow and BC perform on par (BC even leads by 5.3 pp under DINO-WM), and only from N=300 on does Flow lead BC under all world models.

Flow’s diverse candidates thus pay off only when enough of them are sampled. Flow pulls ahead clearly only in specific regimes, namely long horizons and dense replanning. Beyond the choice of prior, the planning setting around it matters more:

  • Reversed comparisons: under LWM, Flow beats CEM by 5.2 pp at N=30 but trails it by 4.8 pp at N=300.

  • Replanning cadence and cost weighting: together they swing success by more than 60 pp, far more than the differences between the priors in our default setting (at most about 13 pp, Table 1).

Finally, priors are efficient regardless of their family: used alone, they cut planning time by 2–15× relative to CEM (Table 10), and Flow+CEM matches the success of the best uninformed planners in 3.6–10× less planning time (Figure 3). We believe this mapping from planning setting to prior is the main practical contribution, and it cannot be predicted from expressivity alone.

W2: Simple environments in which CEM already does well.

We use established benchmarks from latent planning research that span different complexities, including contact-rich manipulation (Push-T, OGBench-Cube) and action chunks of 50–250 dimensions. These are not settings in which CEM already does well: it reaches only 7.8–38.2% on long-horizon Push-T, 46–51% on long-horizon Cube and 59–72% on short-horizon Cube. Priors raise success to up to 67%, 90% and 99%, respectively. Following your recommendation, we performed additional evaluations on the deformable-object manipulation tasks Rope and Granular and found that seeding CEM with a prior gives the lowest mean Chamfer distance on both tasks.

W2 cont.: Recent work [1–3].

We thank the reviewer for these references and will discuss all three in the related work. [1] refines a Diffusion Policy with MPPI, using a learned reward model, a critic and online world-model fine-tuning. It relies on a single prior family, while we compare prior families in an offline, reward-free setting. [3] mainly speeds up each world-model rollout through token dropout. It also shows that action candidates from a pretrained VLA improve planning on long-horizon tasks. This supports the value of priors, but [3] uses one fixed prior and does not study how different priors interact with the planner, which is the focus of our work.
[2] adds an inverse-dynamics consistency term to the CEM cost but keeps an uninformed initialization, so it is complementary to learned priors. We agree that its environments strengthen our study and have extended our evaluation with two of them, Rope and Granular.

Q1: Naming of the deterministic prior.

Thank you for pointing this out. We agree with you and will rename the deterministic prior from "BC” to "Det" in our manuscript to enhance clarity.

Q2: Use a more powerful prior, e.g., a VLA with the world model as verifier.

To clarify, the priors are not trained jointly with the world model: the world models are pretrained and frozen, and the priors are trained afterwards on the same offline data. Additionally, [2] do not seem to use a larger-scale VLA in their experiments. Our framework accepts any proposal distribution, including a VLA. For this study, we deliberately keep architecture, capacity, and training data fixed so that only the distributional form of the prior varies. A large pretrained VLA would differ from our priors in all of these at once. We see VLA priors with a world model as verifier as a promising complementary direction and will add a discussion of them in the related work section. Since some VLAs, such as pi0, use flow-matching action heads, our findings on the Flow prior may also inform such systems.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (1/2)

We thank all reviewers for their constructive feedback. Reviewers found our comparison controlled and fair (uP93, zYY8), the questions we study practically relevant (uP93, zYY8), and the paper clear (mg9Y). In response, we (i) reran Tables 1 and 2 with seven seeds, (ii) added the deformable-object tasks Rope and Granular, (iii) added a controlled data-quality sweep, and (iv) evaluated a stronger, retraining-free autoregressive extension of the short-horizon priors.

Key takeaways

  • Data decides whether a prior can replace search, not the prior family. With goal-directed demonstrations, a prior alone matches or beats CEM at a fraction of the planning time. With undirected data, a prior can still be useful as a seed for CEM.

  • Expressivity matters less than expected. In standard short-horizon planning the three priors are close, and the simplest (deterministic) prior is often as good and the cheapest.

  • Flow pays off in specific regimes: dense replanning, long horizons, and large sample budgets.

  • The planning setting matters more than the choice of prior. Replanning cadence and cost weighting change success far more than switching priors, and the best cost weighting differs per planner.

  • Comparisons can reverse with sample size, so evaluating planners at a single sample size is not enough.

  • Prior+CEM is a robust default. It matches or improves on CEM almost everywhere and needs fewer refinement steps. For longer horizons, retraining the prior at the target horizon works best, but even extending a short-horizon prior autoregressively, without retraining, still outperforms CEM on long-horizon Push-T.

  • Priors are cheap. They train in minutes, and a fraction of the demonstrations (or of goal-directed data) keeps nearly the full benefit.

Table 1 and Table 2 with seven seeds

We reran all experiments behind Table 1 and Table 2 with more seeds (seven instead of three) and report the results below. Almost all success rates stay within a few percentage points of the three-seed values, and none of the conclusions changes: the priors clearly beat uninformed planning on OGBench-Cube and Two-Room, while CEM remains the strongest standalone planner on Wall and both PointMazes, and Flow+CEM and Gaussian+CEM match or improve on CEM in every DINO-WM setting.

Table 1: Priors vs. uninformed planners

World modelEnvironmentFlowGaussianBCCEMGD
LeWorldModelTwo-Room96.00 ± 1.4198.21 ± 0.8699.93 ± 0.1983.93 ± 3.9894.50 ± 1.15
LeWorldModelPush-T83.21 ± 1.9179.71 ± 2.7779.71 ± 4.4886.21 ± 1.7872.86 ± 4.32
LeWorldModelOGBench-Cube99.14 ± 0.5699.86 ± 0.2499.36 ± 0.5669.79 ± 3.0461.00 ± 4.02
PLDMTwo-Room96.57 ± 0.8998.14 ± 1.31100.00 ± 0.0095.29 ± 2.0295.21 ± 1.87
PLDMPush-T83.64 ± 3.4281.07 ± 4.4378.93 ± 3.7978.07 ± 2.6567.50 ± 2.04
PLDMOGBench-Cube98.36 ± 0.8599.79 ± 0.3999.50 ± 0.5859.43 ± 4.5554.50 ± 3.99
DINO-WMWall93.21 ± 1.5888.50 ± 2.0480.29 ± 1.3893.71 ± 1.9191.07 ± 2.21
DINO-WMPush-T76.86 ± 2.3869.64 ± 3.6471.71 ± 3.7373.86 ± 2.8438.57 ± 2.60
DINO-WMPointMaze-UMaze66.64 ± 3.1367.14 ± 3.4560.86 ± 3.1283.29 ± 2.5074.07 ± 3.91
DINO-WMPointMaze-Medium66.43 ± 3.6969.86 ± 3.2068.36 ± 2.9489.07 ± 2.6875.86 ± 3.72
DINO-WM (Straightened)Wall94.50 ± 0.7191.71 ± 1.7581.86 ± 1.6897.57 ± 1.0694.50 ± 1.50
DINO-WM (Straightened)Push-T77.29 ± 3.3872.71 ± 2.9370.93 ± 3.4278.00 ± 4.1143.00 ± 5.66
DINO-WM (Straightened)PointMaze-UMaze80.93 ± 2.3781.36 ± 2.1989.64 ± 3.5894.14 ± 1.2884.86 ± 2.48
DINO-WM (Straightened)PointMaze-Medium68.86 ± 5.5170.29 ± 4.0178.57 ± 3.1784.36 ± 2.6175.79 ± 2.84

Table 2: Priors+CEM

World modelEnvironmentFlow+CEMGaussian+CEMBC+CEMCEM
DINO-WMWall97.43 ± 1.4396.43 ± 0.8994.07 ± 1.5193.71 ± 1.91
DINO-WMPush-T87.64 ± 2.7686.64 ± 2.6185.21 ± 2.8473.86 ± 2.84
DINO-WMPointMaze-UMaze84.93 ± 2.1583.71 ± 3.0175.86 ± 2.9583.29 ± 2.50
DINO-WMPointMaze-Medium89.43 ± 1.5791.14 ± 1.3185.93 ± 1.5189.07 ± 2.68
DINO-WM (Straightened)Wall97.86 ± 1.4697.93 ± 1.3094.14 ± 1.9397.57 ± 1.06
DINO-WM (Straightened)Push-T89.93 ± 2.3289.71 ± 2.0484.86 ± 2.4178.00 ± 4.11
DINO-WM (Straightened)PointMaze-UMaze94.36 ± 1.6894.64 ± 2.1596.43 ± 1.0694.14 ± 1.28
DINO-WM (Straightened)PointMaze-Medium88.50 ± 2.9984.86 ± 1.8989.57 ± 2.3784.36 ± 2.61
AuthorsOpenReview ↗

General response (2/2)

Rope and Granular: CEM vs. Prior+CEM

We additionally performed evaluations on the two DINO-WM deformable tasks Rope and Granular and report the average Chamfer distance (lower is better) over 7 seeds. On both deformable tasks and with both world models, starting CEM from a learned prior gives the lowest mean Chamfer distance. In all but one case Prior+CEM matches or improves on CEM.

TaskWorld modelCEMFlow+CEMGaussian+CEMBC+CEM
RopeDINO-WM0.78 ± 0.290.48 ± 0.150.63 ± 0.180.71 ± 0.30
RopeDINO-WM (Straightened)0.76 ± 0.210.66 ± 0.290.71 ± 0.220.60 ± 0.13
GranularDINO-WM0.31 ± 0.040.27 ± 0.010.26 ± 0.030.18 ± 0.01
GranularDINO-WM (Straightened)0.36 ± 0.020.34 ± 0.030.37 ± 0.040.29 ± 0.02

Autoregressive Extension of short-horizon priors

The table below reports the success rate (averaged over seven seeds) of autoregressively used short-horizon priors. Improvements over Gaussian padding (in percentage points) are reported in the brackets. Extending the short-horizon priors autoregressively improves on Gaussian padding in all 24 world-model/prior configurations. Autoregression bridges a substantial part of the gap between the Gaussian-padding short-horizon priors and the priors trained natively at H=10, but the retrained priors stay ahead in all 12 Prior+CEM configurations, as does the retrained Flow under all four world models. Compared to CEM, the autoregressively extended priors alone outperform it in 11 of 12 configurations, and seeding CEM with them improves on CEM in all configurations, by 9.7 to 19.9 pp.

World modelFlowGaussianBCFlow+CEMGaussian+CEMBC+CEM
LeWorldModel44.79 ± 3.46 (+5.62)45.00 ± 2.52 (+5.83)39.93 ± 1.72 (+1.43)53.64 ± 4.33 (+5.64)57.86 ± 5.21 (+6.86)53.64 ± 3.68 (+2.81)
PLDM39.43 ± 3.31 (+6.60)37.07 ± 4.05 (+3.57)33.50 ± 3.01 (+3.50)53.00 ± 2.57 (+4.17)54.14 ± 4.14 (+1.31)55.07 ± 2.79 (+4.74)
DINO-WM21.57 ± 3.22 (+13.07)18.43 ± 3.25 (+11.10)14.50 ± 1.19 (+7.17)21.50 ± 2.84 (+3.33)28.43 ± 1.90 (+7.26)25.86 ± 2.30 (+6.69)
DINO-WM (Straightened)19.14 ± 2.23 (+14.64)14.64 ± 1.75 (+7.64)11.43 ± 1.51 (+4.93)19.07 ± 1.72 (+3.90)20.07 ± 1.79 (+5.07)17.57 ± 2.30 (+3.57)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.