acceptodds
Under review as a conference paper at ICLR 2027

Fast Planning with Cross-Candidate Inference

Abstract

Sampling-based planning with latent world models typically evaluates every candidate action sequence through a full-horizon rollout. One way to reduce rollout overhead is to fully roll out only a subset of candidates and use the resulting evaluations to reconstruct those of the remaining candidates. The key challenges are to model evaluation correlations among candidates and identify the full-horizon rollouts that are most informative for planning updates. To this end, we propose an efficient planning method for frozen world models that combines cross-candidate inference with rollout allocation guided by action-distribution updates. Using short-horizon predicted states, evaluation trends, and remaining control inputs, the method captures correlations across candidates through a Gaussian process with a trend term. We combine the posterior covariance from this Gaussian process with the local sensitivity of action-distribution updates to derive an update-risk criterion for selecting candidates for full-horizon rollouts. The resulting evaluations are then used to condition the Gaussian process and reconstruct the remaining candidates’ long-horizon evaluations. Observed and reconstructed evaluations then jointly inform the action-distribution update. Experiments across four latent world models and diverse control tasks demonstrate that our method substantially reduces planning latency while maintaining planning performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
World Models, Model-Based Planning, Efficient Inference, Gaussian Processes

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kVWvOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes cross-candidate inference to reduce the cost of sampling-based planning with frozen latent world models. Based on selective full-horizon rollouts, the method uses an update-aware way to select a subset of the candidates for full-horizon continuation, and reconstructs the remaining long-horizon score corrections with a Gaussian process, allowing all candidates to inform the soft action-distribution update. The experiments on DINO-WM and LeWM proves its effectiveness.

Critical strengths
  1. The update-aware objective is a meaningful methodological contribution. The paper shows that score accuracy does not necessarily translate into update accuracy, motivating rollout selection based on the sensitivity of the planning update.
  2. The use of a Gaussian Process with a trend term to reconstruct unobserved score corrections is highly effective. In the evaluated settings, it produces a practical training-free planning wrapper with substantial computational savings.
Critical weaknesses
  1. The method’s robustness and efficiency beyond the evaluated settings remain unclear. Its GP reconstruction and covariance model are tested on a limited set of benchmark families and planning configurations, while larger candidate pools or more frequent replanning may increase computational and memory costs.
  2. The paper reports that its nominal 95% intervals achieve less than 1% empirical coverage. It’s unclear whether the GP covariance is still reliable for ranking candidates, or how posterior update risk relates to realized planning-update error.
Questions for authors
  1. How are GP kernel and trend hyperparameters selected, and are they fixed across tasks, horizons, and world-models?
  2. What happens when the candidate pool is much larger, the action dimension is higher, or the horizon is substantially longer?
  3. Can the method be applied to planners whose update is not a diagonal-Gaussian soft update, such as CEM elite selection or correlated action distributions? Which parts of the derivation would change?
Reviewer_55dqOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method to speed up planning with latent world models by reducing the number of whole rollouts. Instead of evaluating every candidate action sequence to the end, the method uses short-horizon predictions and a Gaussian process to estimate the long-horizon scores of unselected candidates. In this way, it selects candidates for full evaluation based on their expected impact on the planning update, rather than score uncertainty alone.

Critical strengths

The method leverages information from selected rollouts to estimate the remaining candidates' scores. This makes the full candidate pool to contribute to planning. And this approach does not require retraining the world model and demonstrates effectiveness across two different world models and multiple planning tasks.

Critical weaknesses

I think this paper is lack of novelty. The overall idea of selectively evaluating candidates and estimating the remaining scores is closely related to existing surrogate-based optimization and active sampling methods. Also, this method assumes that candidates with similar short-horizon states and remaining-action statistics have correlated long-horizon scores. This assumption may not hold in complex environments with nonlinear dynamics.

Regarding the experiment design, it only uses a short planning horizon, which is 5 steps. It remains unclear whether the method is effective for longer horizons, where prediction errors and trajectory divergence may increase. Also, the gaussian process inference and greedy candidate selection introduce additional computational overhead, which may limit efficiency gains for larger candidate pools.

A minor point is that the method primarily aims to approximate the original full-rollout planner more efficiently, rather than improving planning quality or addressing inaccuracies in the underlying world model.

Questions for authors

See weaknesses

Reviewer_8bFpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Planning with a learned simulator by evaluating different action sequences into the long-term future could be computationally expensive. The paper proposes to briefly simulate all candidate sequences but only fully simulate a selected few. It then uses similarities between candidates to estimate how the others would perform. It decides which candidates to complete by using their expected influence on the planner's next update. The paper reports empirical results with less predictions steps and improved computation with similar average success rate to full evaluation.

Critical strengths

The method targets an important problem and it can be used with an existing model without retraining. The experiments demonstrate certain level of benefits of the proposed method.

Critical weaknesses

First, I think it misses a highly related work that focus on carefully judging which actions to explore, such as Bayesian optimized Monte Carlo planning (John Mern et al.). Though that setting is different, the authors should do more literature survey and include relevant discussions.

Second, the uncertainty model is severely miscalibrated in the reported diagnostic. Appendix B reports less than 1% coverage for intervals intended to contain the true values 95% of the time, but the method uses these uncertaity estimates to choose which candidates to simulate fully and also to lower the estimated scores of uncertain candidates, hence inaccurate estimates can affect planning.

Third, there are important empirically results contradict one another. figure 6 and table 10 reports different success rate under the setting H=10, B=60 and prefix length h=4. figure 5 and table 12 also disagree about which allocation method achieves the smallest update error. These discrepancies would affect central conclusions and not merely a presentation issue.

Fourth, there is a nontrivial theoretical gap between risk reduction and implemented planning update: eq 11 uses local error for posterior-mean reconstruction, but Eq15 actually uses uncertainty-panelized estimates which could introduce an additional bias term omitted from the criterion.

Last, the candidate representation does not preserve the action order information. Section 3.1 summarizes remaining actions using mean and covariance information, and these cannot distinguish reordered action sequences. That said, plans with the same prefix but differently ordered remaining actions can therefore receive identical features despite very different outcomes. This appears to be a serious problem and it is not clear from the paper why/when this would not hurt.

Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.