acceptodds
Under review as a conference paper at ICLR 2027

How Much Planning Is Enough? Reducing Search and Computation in World-Model Planning

Abstract

Visual world models enable goal-directed control through decision-time action search, but their deployment efficiency is often limited by conservatively large planning budgets. We show that competitive task performance can be achieved without agreement with the Full-budget action, that sufficient budgets vary across model–task pairs, and that iterative planners repeatedly encode solve-invariant context. To address these inefficiencies, we propose SufficientPlan, a simple deployment framework that requires no modification to pretrained world models or planner updates. Its Paired Sequential Budget Certification (PSBC) component uses paired closed-loop evidence to search for and certify a reduced model–task-specific budget within a predefined Full-performance tolerance. Its StaticContext Reuse (SCR) component caches observation and goal representations across search iterations while preserving candidate-dependent planning and selected actions. Experiments across multiple world-model backbones and visualcontrol tasks show that SufficientPlan substantially reduces search budgets and planning latency while maintaining competitive control performance.

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

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_B9MVOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the SufficientPlan framework, which improves the computational efficiency of WM-based planning by selecting a sufficient planning budget using PSBC and reducing redundant encoding operations using SCR.

Critical strengths

A notable advantage of the framework is that it can be applied to existing world models without requiring retraining. The experimental results across a variety of models and tasks show meaningful reductions in computational cost while preserving task success rates.

Critical weaknesses
  1. While the framework effectively reduces planning costs at deployment time, the calibration process itself may introduce additional computational overhead. I would be interested to know whether the overall approach remains efficient once the calibration cost is taken into account, particularly in settings with a relatively small number of deployments. In addition, the framework evaluates candidate budgets sequentially using fixed increments. Could the authors clarify the motivation behind this design choice? A comparison of calibration costs, as well as a discussion of alternative search strategies, would help clarify the overall efficiency of the proposed method.

  2. The framework assigns a fixed planning budget to each model–task pair. However, planning difficulty may vary considerably across states within the same task. For instance, different object configurations or environmental conditions may require more extensive search. In such cases, a fixed budget could lead to unnecessary computation in easier states, or to failures in more challenging ones. Could the authors comment on the robustness of the method under such variation? An analysis of state-dependent planning difficulty, or a comparison with adaptive planning methods, would help address this concern.

Questions for authors

The main questions addressed in the Critical Weaknesses above.

7 comments

As the rebuttal period draws to a close, we hope that our responses have effectively addressed your concerns. Thank you again for your time and consideration.

I appreciate the authors' detailed responses and additional experiments. However, my main concern regarding the use of fixed planning budgets remains insufficiently addressed.

While the additional distance-based analysis provides some evidence of robustness across different initial configurations, it does not fully capture state-dependent planning difficulty. In particular, without a comparison with adaptive planning methods, it remains unclear whether the proposed fixed-budget approach provides a reasonable trade-off between computational efficiency and task performance across different states.

Therefore, I increase my score. However, as my main concern has not been fully resolved, my assessment still leans toward rejection.

Thank you for acknowledging our additional analyses and for raising your score. To directly address your remaining concern, we have added a comparison with online state-adaptive planning in Section C.9, “Comparison with State-Adaptive Planning,” of the newly revised manuscript.

We evaluate Full, PSBC, and three compute configurations of an online adaptive baseline on 500 matched held-out episodes per method on INTACT/Push-T. The adaptive baseline monitors improvement in elite candidate costs and stops each planning solve accordingly, allowing search depth to vary across states. Its thresholds are selected on development episodes and frozen before testing. All methods use the same pretrained model, candidate population, planning horizon, and evaluation episodes, with SCR disabled.

PSBC achieves the same observed success rate as Full and the high-compute adaptive configuration (93.0%), while reducing planning time by 72.9% and 69.1%, respectively. Compared with the medium-compute adaptive configuration, PSBC uses 28.2% less planning time and achieves a 1.4-percentage-point higher observed success rate. More aggressive adaptive stopping reduces computation further, but success decreases to 90.0%.

This comparison directly supports the practical value of a calibrated fixed budget in the tested setting. Although planning requirements vary across states, online adaptation does not automatically produce a better computation–performance trade-off: PSBC retains Full’s observed success rate at substantially lower cost, while the lowest-cost adaptive configuration trades some task success for further savings.

The newly added table is reproduced below. SR intervals are Wilson 95% confidence intervals. Time is warmed, GPU-synchronized, batch-amortized planning time on RTX A5000 GPUs, including adaptive stopping overhead. Adaptive denotes our implemented online elite-cost-stall baseline.

MethodHeld-out SR, % [95% CI]Mean candidate evaluations/episodeMean planning time (s/episode)Iterations/solve: mean [P10, P90]
Full93.00 [90.42, 94.92]9,882.01.42030.00 [30, 30]
PSBC (2,400)93.00 [90.42, 94.92]2,683.20.3848.00 [8, 8]
Adaptive (low compute)90.00 [87.06, 92.33]1,366.20.2003.90 [3, 7]
Adaptive (medium compute)91.60 [88.84, 93.73]3,735.00.53611.24 [3, 22]
Adaptive (high compute)93.00 [90.42, 94.92]8,711.41.24626.69 [16, 30]
Reviewer_E7yjOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SufficientPlan, a framework for reducing the deployment cost of visual world-model planning without retraining the underlying models or changing their native optimization rules. It combines two components: Paired Sequential Budget Certification (PSBC), which selects a fixed, model–task-specific planning budget, and Static-Context Reuse (SCR), which caches observation and goal representations. Across five backbones and four tasks, covering 16 model–task pairs, PSBC reduces candidate evaluations by 70% on average, with 14 configurations remaining within two percentage points of Full-budget success. Matched-budget SCR experiments report speedups of up to 2.18 without changing evaluated episode outcomes.

Critical strengths
  1. The paper addresses a useful practical problem with a clear separation of mechanisms.
  2. The experiment looks reasonably broad, and the results show the method's effectiveness.
Critical weaknesses
  1. I think the experiments establish savings relative to Full planning more clearly than the added value of PSBC over simpler budget selection. It remains unclear how much of the reported benefit requires PSBC, as opposed to simply tuning the iteration budget on a development set. Extending this comparison to representative easy and refinement-sensitive model–task pairs, using identical budget grids and calibration data, would substantially strengthen the contribution.
  2. Table B shows two PRISM/Push-T budgets accepted after 20 episodes with posterior probability 0.956, whose probabilities fall to 0.510 and 0.799 when all 100 calibration outcomes are considered. This questions the reliability of early budget certification.
Questions for authors
  1. How sensitive are the selected budgets and subsequent held-out degradation to the calibration split and episode order, particularly for certificates issued after only 20 episodes? Would a minimum-sample requirement or a small independent confirmation stage improve stability at an acceptable additional calibration cost?
  2. About the related work, maybe the authors should discuss statistical algorithm configuration and post-hoc calibration. For example, Learn then Test calibrates predictive algorithms to satisfy explicit risk requirements without retraining [1]. (It does not affect this paper's contribution on world model planning, just a suggestion). [1] Learn then Test: Calibrating Predictive Algorithms to Achieve Risk Control
Reviewer_LatTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SufficientPlan to accelerate world-model planning through paired sequential calibration of search budgets and reuse of observation and goal encodings. Paired sequential calibration selects a reduced budget using Bayesian comparisons of paired task outcomes, while caching avoids repeated encoding within each planning solve. Across 16 model–task pairs, calibrated planning uses 30% of the full budget on average, with success-rate degradation within 2% in 14 pairs. Further timing and numerical equivalence tests support behavior-preserving caching, with speedups up to 2.45×.

Critical strengths
  • SufficientPlan combines outcome-based budget calibration with reuse of invariant encodings, addressing both excessive search and redundant computation without retraining world models.
  • PSBC evaluates whether reduced search preserves closed-loop task success, rather than whether it reproduces full-budget actions. This allows computational savings even when action sequences differ substantially.
Critical weaknesses
  • The reliability of sequential certification remains insufficiently validated. Proposition 1 provides limited justification for the reliability of the overall budget-selection procedure. Table B shows that two budgets certified at a posterior probability of 0.956 fall to 0.510 and 0.799 after additional observations. Although this does not invalidate the Bayesian formulation, it highlights the need to assess certification stability and false-certification rates across repeated calibration runs.

  • The methodological advantage over ordinary budget tuning is unclear. PSBC ultimately selects a fixed budget for each model–task pair, making its practical role similar to offline hyperparameter tuning. The main experiments demonstrate redundancy in full-budget planning, but do not establish that sequential certification yields better held-out performance–compute trade-offs than simpler selection methods under matched calibration costs. The single-case comparison in Appendix D is insufficient to establish this advantage, particularly since the SCR component is a straightforward caching optimization.

Questions for authors
  • What does PSBC gain over simpler budget tuning under matched calibration costs? Can the authors compare against empirical and fixed-sample posterior selection across model–task pairs, reporting held-out performance, selected budgets, and total calibration cost?
1 comment

As the rebuttal period draws to a close, we hope that our responses have effectively addressed your concerns. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.