PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Changbai Li, Sirui Li, Yichen Yang, Tongfei Chen, Zichao Feng, Shuwei Shao, Huobin Tan

arXiv:2610.08350v1cs.ROcs.AI

TL;DR

World-model planning can incur substantial decision-time computation from large fixed search budgets and repeated encoding of unchanged context. SufficientPlan certifies reduced model–task-specific budgets and reuses static representations; experiments report lower search computation and latency with competitive control performance.

  • Problem

    World-model planners commonly use large, globally fixed search budgets, although additional search may not improve closed-loop task performance and the sufficient budget varies across model–task pairs.

    Source: §1 · p. 2

  • Method

    SufficientPlan combines paired closed-loop budget certification within a predefined tolerance with reuse of observation and goal representations across search iterations, without modifying the pretrained model or planner updates.

    Source: §1 · p. 2

  • Results

    Across multiple backbones, planners, and visual-control tasks, SufficientPlan substantially reduces candidate evaluations and planning latency while retaining competitive control performance.

    Source: §6 · pp. 9–10

  • Takeaways & Limitations

    Outcome sufficiency and computation reuse are complementary principles for efficient world-model planning.

    Source: §6 · p. 10

    Proposition 1 does not imply simultaneous frequentist coverage over all budgets and inspection times or identical empirical success rates on every finite held-out sample.

    Source: §4 · p. 7

Abstract

from arXiv · show

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 {Static-Context 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 visual-control tasks show that SufficientPlan substantially reduces search budgets and planning latency while maintaining competitive control performance.

1 INTRODUCTION

Fixed search budgets can spend computation beyond what closed-loop performance requires, while planners also repeat encodings of unchanged context. SufficientPlan addresses both inefficiencies at deployment time through outcome-based budget certification and static-context reuse.

1 INTRODUCTION

  • On Push-T, calibrated and Full planners achieved 96% success with 2,400 and 9,000 scored sequences, respectively, despite following distinct trajectories.This illustrates that comparable task performance need not require matching the Full-budget action.Source: §1 · p. 2
  • PSBC uses paired development episodes to certify a reduced model–task-specific budget within a predefined performance tolerance, then freezes it for held-out deployment.It evaluates task outcomes rather than action agreement with Full planning.Source: §1 · p. 2
  • SCR reuses observation-history and goal encodings across search iterations, reducing solve-static encoding operations from 2K to 2 over K iterations.Candidate-dependent planning, optimizer updates, selected actions, and closed-loop behavior remain unchanged.Source: §1 · p. 2
  • SufficientPlan reduces search computation and planning latency while preserving competitive closed-loop performance across world-model backbones and control benchmarks.It combines outcome-sufficient budget certification with reuse of solve-invariant representations.Source: §1 · p. 2

2 RELATED WORK

Prior work develops predictive latent representations and improves planning efficiency through policies, priors, or sampling refinements. SufficientPlan instead keeps the pretrained model and planner fixed while reducing decision-time computation.

2 RELATED WORK

  • Latent world models support planning or policy optimization in compact representations, including methods that do not reconstruct pixel-perfect futures.The surveyed approaches include PlaNet, Dreamer, MuZero, EfficientZero, and TD-MPC.Source: §2 · p. 2
  • PSBC compares budget proposals with the Full-budget reference on paired calibration episodes and freezes the first proposal meeting its certification criterion for deployment.Uncertain proposals receive more paired evidence; insufficient proposals are increased.Source: Figure 2 · p. 3
  • Unlike methods that primarily improve representations, dynamics, or action priors, SufficientPlan keeps the pretrained model and planner fixed and reduces decision-time computation.PSBC selects a model–task-specific budget for held-out deployment, while SCR reuses solve-invariant encodings within search.Source: §2 · p. 3
  • Prior planning-efficiency methods use learned policies or behavior priors, sample-efficient CEM refinements, or less frequent replanning.Sampling-based optimizer costs grow with candidate population and search depth.Source: §2 · p. 3

3 MOTIVATION AND ANALYSIS

The analysis identifies excess search beyond outcome sufficiency and repeated encoding of solve-static context as distinct sources of planning overcomputation. Their significance varies by model–task pair, and reuse reduces measured latency while preserving evaluated outcomes.

3 MOTIVATION AND ANALYSIS

  • On INTACT Push-T, calibrated planning uses 2,400 scored sequences versus 9,000 for Full planning, with both achieving 96% development-set success.The planners follow distinct trajectories despite equal aggregate success, motivating paired task-outcome assessment rather than action matching.Source: §3 · p. 3
  • Certified budgets range from 3.3% to 80.0% of Full planning across evaluated model–task pairs, so a uniformly small or large budget may misallocate computation.INTACT on Cube uses 3.3%, whereas LeWM on Reacher uses 80.0%.Source: §3 · p. 4
  • Across K iterations, planners encode two solve-static inputs repeatedly for 2K calls, motivating reuse alongside reduced candidate evaluation.The two sources are search beyond outcome sufficiency and repeated solve-static encoding; SCR illustrates reuse of the observation history and goal.Source: §3 · p. 4, Figure 3 · p. 4
  • At 30 iterations, SCR reaches a 2.45× speedup on Fast-LeWM/Push-T, while matched-budget evaluations preserve episode-success vectors with zero outcome flips.At one iteration, repeated timing measurements show no reliable speedup.Source: §3 · p. 4

4 METHOD

SufficientPlan reduces planning cost in two ways: PSBC certifies a lower model–task-specific search budget from paired closed-loop outcomes, and SCR reuses solve-invariant representations during each retained search. Both operate at deployment time without retraining the model or changing the planner’s objective or updates.

Paired Sequential Budget Certification

  • PSBC compares candidate and Full budgets on the same development episodes and certifies a proposal when its posterior probability of staying within the allowed performance tolerance meets the acceptance threshold.Uncertain proposals receive more paired episodes; if no reduced budget is certified, PSBC falls back to Full planning.Source: §4 · pp. 4–6

Posterior certification guarantee

  • PSBC’s certification guarantee is posterior under the specified paired-outcome model, not simultaneous frequentist coverage or identical success rates on every finite held-out sample.Source: §4 · p. 7

Budget proposal and deployment

  • PSBC returns the first sufficient proposal in a predefined lower-budget sequence, defining a certified operating point rather than a globally minimal budget.The procedure does not assume closed-loop performance improves monotonically with increasing budget.Source: §4 · p. 7

Static-Context Reuse

  • SCR moves observation-history and goal encoding outside the refinement loop because those inputs remain fixed while candidate action populations change.Its solve-local cache is discarded before the next environment step; candidate-dependent sampling, encoding, rollout, cost evaluation, and optimizer updates still run for each population.Source: §4 · p. 7

Computation reduction and planning equivalence

  • Across K refinement iterations, SCR reduces solve-static encoder operations from 2K to two, while preserving fixed-budget control behavior under the stated deterministic evaluation conditions.The reported encoder-call reduction measures neither total planner FLOPs nor latency.Source: §4 · pp. 7–8

5 EXPERIMENTS

Across the evaluated model–task pairs, PSBC substantially reduces search budgets while usually keeping held-out performance close to Full planning; SCR preserves evaluated outcomes and reduces latency. The experiments also examine calibration behavior and the incremental effects of adding each component.

5.1 EXPERIMENTAL SETUP

  • The evaluation covers five frozen world-model backbones, four visual-control tasks, and their native planners, with budgets calibrated on development episodes and frozen for held-out evaluation.It reports success rate, budget relative to Full planning, and matched-budget latency, speedup, and outcome flips for SCR.Source: §5.1 · p. 8

5.2 BUDGET CERTIFICATION

  • PSBC uses 30.0% of the corresponding Full budget on average across 16 backbone–task pairs, and 14 configurations finish no more than 2 percentage points below Full planning.Table 1 compares Full and PSBC-calibrated planning across world-model planners and control tasks.Source: §5.2 · p. 9
  • Certified budgets range from 3.3% to 80.0% of Full planning; PSBC selects 13.3% for LeWM on Cube and 80.0% for Reacher.Held-out decreases of 6 percentage points on Fast-LeWM/Push-T and 3 points on C-JEPA/Push-T are reported without recalibration.Source: §5.2 · p. 9

5.3 STATIC-CONTEXT REUSE

  • With matched budgets and evaluation conditions, SCR preserves episode-success vectors with zero outcome flips and achieves up to 2.18× planning speedup across evaluated settings.The gain varies with the runtime share spent on repeated static encoding.Source: §5.3 · p. 9
  • On Fast-LeWM/Push-T, the static-encoder-call reduction rises from 0% at one iteration to 96.7% at 30, while measured speedup reaches 2.45× at 30 iterations.Repeated measurements show no reliable speedup at one iteration.Source: §5.3 · p. 9

5.4 FURTHER ANALYSIS OF PSBC

  • In the PRISM/Cube case study, paired posterior certification rejects a budget accepted by empirical success rates alone, while stricter settings require more evidence or leave a proposal uncertified within the available horizon.Early rejection reduces candidate-side calibration evaluations.Source: §5.4 · p. 9

5.5 ABLATION STUDY

  • On the frozen Push-T development set, PSBC reduces INTACT’s search budget from 9,000 to 2,400 scored sequences while preserving 96% success; adding SCR leaves selected budgets and outcomes unchanged.For Fast-LeWM, PSBC reduces the budget to 26.7% of Full planning with a 1-point success decrease; SCR’s latency benefits are reported separately.Source: §5.5 · p. 9

6 CONCLUSION

The conclusion identifies different actions under reduced and Full budgets, despite equal aggregate task performance, and repeated encoding of solve-invariant observations and goals as sources of overcomputation. SufficientPlan addresses these with budget certification and representation reuse, reducing candidate evaluations and latency while retaining competitive control performance.

6 CONCLUSION

  • World-model planners can produce different actions under reduced and Full budgets despite achieving the same aggregate task performance.Source: §6 · p. 9
  • SufficientPlan requires neither world-model retraining nor changes to the native planner, and combines paired-evidence budget certification with reuse of solve-static representations.Source: §6 · p. 9
  • Across multiple backbones, planners, and visual-control tasks, SufficientPlan reduces candidate evaluations and planning latency while retaining competitive control performance.The conclusion identifies outcome sufficiency and computation reuse as complementary principles for efficient world-model planning.Source: §6 · p. 10

AI USE STATEMENT

The authors report using AI tools to polish language and improve readability, while taking full responsibility for the paper and its results.

AI USE STATEMENT

  • AI tools were used to polish language and improve readability; the authors take full responsibility for the paper and its results.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

The authors report simulated visual-control experiments, with no human-subject experiments or personal-data collection.

ETHICS STATEMENT

  • The work uses simulated visual-control benchmarks, involves no human-subject experiments or personal-data collection, and identifies no anticipated direct ethical risks from the reported simulations.Physical deployment should follow the safety requirements of its intended application.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

The paper describes its experimental protocols, model backbones, and planned reproducibility materials. Evaluations use distinct development and held-out episodes, with paired comparisons on the same episodes, states, and goals.

REPRODUCIBILITY STATEMENT

  • The reproducibility statement says the paper and supplement provide model, training, dataset, and evaluation details, and that implementation and configuration files are planned for release.Reported experiments use consistent settings unless otherwise specified.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 DATASETS AND EVALUATION PROTOCOL

  • Evaluations cover Push-T, Cube, Reacher, and TwoRoom, using frozen development splits for calibration and episode-disjoint held-out splits for final evaluation.Paired candidate-budget and Full comparisons use the same episodes, initial states, and goals.Source: §A.1 · p. 13

A.2 BACKBONES AND PRETRAINED CHECKPOINTS

  • The study evaluates LeWM, Fast-LeWM, PRISM, C-JEPA, and INTACT, using released checkpoints when available and otherwise training under official recommended configurations.All checkpoints remain frozen during calibration and held-out evaluation; the implementation and configuration files are planned for release.Source: REPRODUCIBILITY STATEMENT · p. 10, §A.2 · p. 13
  • Figure A compares global fixed-budget allocation with model–task-specific PSBC allocations across the evaluated pairs.The figure describes the counts of under-, matched-, or overallocated pairs and selected budget fractions.Source: §A.2 · p. 14
  • 34.28 s for Full planning versus 10.93 s for PSBC yields a 23.35 s per-episode time saving on PRISM/Cube.The deployment budgets are 9,000 for Full planning and 600 for PSBC.Source: §A.2 · p. 14

A.3 PLANNING AND EVALUATION SETTINGS

The section specifies planning-budget and evaluation settings, then details PSBC certification and SCR’s equivalence and latency properties. It also outlines additional analyses of budget variation, calibration stability, deployment efficiency, and SCR.

A.3 PLANNING AND EVALUATION SETTINGS

  • The planning budget is the number of candidate action sequences scored per solve; fixed-population CEM uses 300 candidates per iteration and varies refinement iterations.The allowed PSBC success-rate degradation is two percentage points, and its tolerance applies to posterior population-level degradation under the development distribution, not every finite held-out sample.Source: §A.3 · pp. 14–15

B.1 PSBC POSTERIOR CONSTRUCTION AND SEQUENTIAL DECISION

  • PSBC tests candidate budgets in ascending order using paired outcomes, returns the first proposal certified within tolerance, and falls back to Full if none is certified within available development episodes.Certification is posterior-based and is not a simultaneous frequentist coverage guarantee or a claim of global budget minimality.Source: §A.3 · p. 15, §B.1 · p. 17

B.2 EXACT PLANNING EQUIVALENCE OF SCR

  • Under the stated deterministic and matched-state assumptions, SCR and the original planner produce identical candidate populations, costs, optimizer states, and selected actions at every refinement iteration.The theorem is stated in exact arithmetic; finite-precision equality additionally requires matched operation ordering, precision, and deterministic kernels.Source: §B.2 · pp. 17–18
  • On INTACT Push-T, across 100 development episodes, 8 and 30 CEM iterations achieve the same 96% success rate despite selecting different actions.The figure also reports mean RMS distance between the first action block at each iteration and its 30-iteration counterpart, with paired bootstrap 95% confidence intervals.Source: Figure C · p. 18

B.3 COMPUTATION AND LATENCY ANALYSIS OF SCR

  • SCR reduces solve-static encoder calls by 1−1/K, but leaves candidate-dependent computation unchanged, so latency gains depend on the share of time spent on static encoding.The analysis says benefits increase with refinement depth, while candidate-dependent computation bounds asymptotic speedup.Source: §B.3 · p. 19

C ADDITIONAL EXPERIMENTS

  • The additional experiments examine model–task budget variation, PSBC stability across seeds and deployment efficiency, INTACT deployment strategies, and SCR scaling and numerical equivalence.The certification trace on PRISM/Cube shows the 600-sequence proposal crossing the 0.95 threshold after 100 paired calibration episodes.Source: Figure D · p. 19, C ADDITIONAL EXPERIMENTS · p. 20

C.1 ADDITIONAL ANALYSIS OF PSBC

These analyses show that calibration behavior and useful planning budgets vary across model–task pairs, while reduced budgets can preserve task success. They also examine deployment savings, SCR’s execution equivalence, and how paired evidence affects PSBC certification.

C.1 ADDITIONAL ANALYSIS OF PSBC

  • A 26.7% global budget falls below the selected budget for 6 pairs, while 80.0% exceeds it for 14 of 16 pairs.Success–budget profiles vary across backbones and tasks; multi-seed evaluation generally finds PSBC success comparable to Full planning, with small variation relative to backbone–task differences.Source: §C.1 · pp. 20–21
  • Eight and 30 CEM iterations both achieve 96% success, although their first action blocks differ in all 100 initial solves.In a separate calibration example, the 600-sequence candidate crosses the 0.95 certification threshold after 100 paired episodes; posterior probabilities need not change monotonically.Source: §C.1 · p. 20
  • 34.28 to 10.93 seconds per episode: PSBC reduces average PRISM/Cube planning time by 23.35 seconds per episode.The passage attributes the lower deployment-time planning cost to the reduced search budget.Source: §C.1 · p. 20

C.2 INTACT DEPLOYMENT ABLATION

  • Direct prediction has the lowest search cost but the lowest success rate; actor-guided initialization, PSBC, and SCR each contribute distinct deployment comparisons.SCR removes repeated solve-static computation without changing the selected budget or success outcome.Source: §C.2 · p. 21

C.3 ADDITIONAL ANALYSIS OF SCR

  • At 30 iterations, SCR removes 96.7% of solve-static encoder calls and reduces planning latency from 11.4736 to 4.6796 seconds, a 2.45× speedup.At one iteration, no cross-iteration reuse is available and the measured timing difference is not reliable.Source: §C.3 · p. 21
  • At the 9,000-sequence budget, SCR preserves candidate populations, costs, CEM updates, selected actions, solver random-number states, model buffers, and scored-sequence counts.The reported numerical-equivalence contract covers both INTACT and Fast-LeWM.Source: §C.3 · pp. 21–22

D.1 COMPARISON WITH SIMPLER SELECTION CRITERIA

  • The 300-sequence candidate meets the empirical tolerance boundary but has only 0.503 posterior non-inferiority probability; the 600- and 2,400-sequence candidates reach 0.981 and 0.984.The 1,200-sequence candidate’s probability is 0.911, below the 0.95 threshold despite a success rate 4 percentage points above Full; PSBC does not assume monotonic performance with budget.Source: §D.1 · p. 23

D.2 SEQUENTIAL INSPECTION EFFICIENCY

  • Sequential inspection rejects the 300-sequence proposal after 20 paired episodes and certifies 600 after 100, avoiding 80 evaluations of the rejected proposal.Retrospective checks find that the two Push-T certificates fall below threshold after unused outcomes are inspected, while the two Cube certificates remain above it.Source: §D.2 · pp. 23–24, §D.3 · p. 24

D.4 SENSITIVITY OF THE CERTIFIED PROPOSAL

The evidence needed to certify the 600-sequence proposal varies with the tolerance and posterior threshold. SufficientPlan’s calibration and reuse benefits also have scope boundaries tied to distribution changes and the share of runtime spent on solve-static computation.

D.4 SENSITIVITY OF THE CERTIFIED PROPOSAL

  • With δ = 2 and τ = 0.95, the 600-sequence proposal is certified after 100 paired episodes; a 5-point tolerance permits certification after 80 for τ = 0.90 or 0.95.A threshold of 0.99 is not reached under the default tolerance, and under zero-point tolerance certification occurs only for τ = 0.90.Source: §D.4 · p. 24
  • The mechanism analysis is a case study, not a cross-task claim about calibration savings; cross-backbone and cross-task held-out deployment results are reported separately.If the 600-sequence proposal is not certified, this analysis does not infer a final deployment budget without evaluating the remaining proposals.Source: §D.4 · p. 24

E LIMITATIONS

  • PSBC calibrates on a frozen development distribution, so substantial changes in the model, task dynamics, or goal distribution may require recalibration.SCR’s speedup depends on solve-static computation’s share of runtime and may be limited when candidate-dependent rollout and cost evaluation dominate.Source: E LIMITATIONS · p. 25

Related papers

Loading 2610.08350v1…