acceptodds
Under review as a conference paper at ICLR 2027

StopTheDream: Outcome-Grounded Candidate Comparison for Adaptive Search in World-Model Planning

Abstract

A world-model planner can always select the best candidate it has imagined, even when its predictions provide weak evidence that the selected action will achieve the goal. Generating more candidates can help, but adds computation and can amplify the same scoring error. We introduce StopTheDream, a framework that connects candidate actions to their predicted consequences and uses this evidence for candidate comparison and bounded search. The procedure makes generation, comparison, and computation allocation explicit, grounding its scoring objectives in observed actions and outcomes. On 120 hierarchical world model (HWM) maze tasks, the policy solves 117 tasks versus CoVO-MPC’s 113, using 48.58% fewer latent transitions and 24.38% less planning time. The original HWM planner solves 105 tasks with the same world model. On 1,000 CAST visual trajectories, it attains 1.3629 absolute trajectory error, 4.18% below a locally trained PiJEPA baseline, with 54.11% fewer predictor calls. These results show higher observed planning quality with substantially less predictive computation, supporting joint evaluation of decision quality and computational cost

Primary area:
applications to robotics, autonomy, planning
Keywords:
world models, adaptive computation, value of computation, action-conditioned prediction, embodied decision-making

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

This paper studies candidate selection and computation allocation in world-model planning. The authors argue that selecting the best available candidate does not establish that it is good enough to execute or determine whether further search is worthwhile. StopTheDream combines candidate generation, outcome-supervised comparison, and bounded additional search. In the main HWM implementation, a controller first chooses between one full-budget search and two smaller searches. A scorer then estimates execution success from candidate actions and predicted trajectories. Low candidate scores trigger one additional search. The paper evaluates quality–computation trade-offs across planning, feedback scheduling, and active information acquisition.

Critical strengths
  1. The paper highlights that selecting the best available candidate does not establish whether it is worth executing or whether further computation is justified. This distinction addresses a practical challenge in budget-constrained planning.

  2. Supervising candidate evaluation with realized outcomes connects planning more directly to control performance. The integration of candidate selection and computation allocation produces favorable quality–computation trade-offs on HWM and CAST.

Critical weaknesses
  1. In the HWM implementation, low scores trigger one additional search. However, the paper does not sufficiently explain when this identifies a recoverable search failure rather than a state where further search is unlikely to help. The threshold study shows the overall effect of changing the rule, but leaves this distinction unresolved.

  2. HWM uses candidate scoring and additional search, while other applications involve feedback scheduling, observation reuse, or task-type routing. Key implementation choices are spread across task-specific appendices, making it difficult to distinguish the common algorithm from separate application designs. Figure 3 also leaves the additional-search condition implicit and labels supervision targets as “Allocator loss,” adding to the reading difficulty.

Questions for authors
  1. What evidence supports using low candidate scores to trigger another search? Are the gains from additional search concentrated in these cases?

  2. Disabling additional search still solves 116/120 HWM tasks, compared with 117/120 for the full method. Since this variant retains the initial allocator, how much of the improvement in success and computation comes from initial allocation, candidate scoring, and additional search, respectively?

  3. Which components are shared across the evaluated tasks, and which are task-specific? Please make this distinction explicit in the method description.

7 comments
AuthorsOpenReview ↗

Response to Reviewer ftBs (1/2)

Dear Reviewer ftBs,

Thank you for your careful reading and precise questions, and for finding that supervising candidate evaluation with realized outcomes connects planning more directly to control performance.

Question 1 and Weakness 1: Evidence for searching again on low scores, where the gains concentrate, and when a failure is recoverable

In our framework, a low score for every plan indicates that the current plans are likely to fail, and two results support using it as the trigger. First, plans chosen when both scores are below 0.5 fail three times as often in re-runs (6 of 33 runs vs. 16 of 264). Second, searching again only on these 11 low-score tasks solves 117 tasks, against 116 with no extra search and 115 when searching again after all 99 two-plan tasks. The score indicates that a decision is at risk; whether the failure is recoverable depends on its cause. A failure is recoverable when the first two searches, each with half the samples, missed a good route by chance: the additional full search draws new samples and can find it (Figure 4b, c). A failure is not recoverable when the world model misjudges a whole region; the rule then costs only one extra search per task.

Another search can then help, because each of the first two searches used half the samples and can miss a good route by chance. The additional search is a full search with new samples, twice as many as each smaller search, so the additional search can cover more routes (Figure 4b, c shows one case). When the world model misjudges a whole region, another search cannot help, and the rule then costs only one search per task.

Notably, re-runs of the chosen plans support the design: plans with low scores fail three times as often. We took the 99 HWM test tasks where the allocator chose two half-size searches. From the same saved state, we executed the plan chosen without the extra search three times, with different seeds (Section 5.5, Appendix K.2):

Scores of the two plansTasksFailed runs
Both below 0.5 (the additional search runs)116 of 33 (18%)
At least one at 0.5 or above8816 of 264 (6%)

Yes, the gains from additional search are concentrated in the low-score tasks (Table 2; threshold fixed on development data; one latent transition is one predicted step of one rollout):

When the planner searches once moreExtra searchesLatent transitionsTasks solved
Never0383.3 million116
Only when both plans score below 0.5 (StopTheDream)11384.4 million117
After every task with two plans99394.0 million115

Searching again after every task solves fewer tasks. When a plan with a high score already exists, another search rarely adds a working route. Each extra plan also gives the scorer one more chance to pick a plan that the scorer overrates. In the low-score tasks, the additional search is useful: the scorer preferred the new plan in 7 of the 11 tasks (Table 2).

CAST shows the same pattern: the less confident the scorer, the more often more proposals contain a better one. We grouped the 1,000 CAST trajectories by the confidence of the scorer after two proposals. The other two proposals contain a better one in 68% of the 200 least confident trajectories, against 57% of the 200 most confident (Appendix K.3).

AuthorsOpenReview ↗

Response to Reviewer ftBs (2/2)

Question 2: Contribution of initial allocation, candidate scoring, and additional search

In our framework, additional search contributes one solved task (116 → 117) for 0.3% more computation, and allocation and scoring together contribute the rest of the success gain (116 vs. 105 for the original HWM planner). The computation saving comes from none of the three: it comes from replanning with half the usual samples after the first four actions, while the three components add almost no computation. Within the first decision, scoring improves the choice between the two plans: on 40 held-out states where both plans were executed, the scorer's choice succeeds in 38, against 35 for the planner's own cost. We measured each part on the same 120 HWM tasks and seeds, and ran two other planners at about the same computation (Tables 1 and 2, Section 5.2, Appendix K.2):

RunTasks solvedLatent transitions
StopTheDream: allocation, scoring, additional search117384.4 million
StopTheDream without additional search116383.3 million
CoVO-MPC with half budget, without allocation and scoring106376.8 million
Smooth MPPI with half budget, without allocation and scoring104371.0 million
Original HWM planner with full budget, without allocation and scoring105741.7 million

Most of the success gain comes from the decision at the start of each task. Planners with the same half budget but without allocation and scoring solve more than ten fewer tasks. The additional search then adds one more task at almost no extra computation.

Scoring alone already chooses better than the planner's own cost. The planner's own cost is the predicted distance between the final state of a plan and the goal. On HWM, we executed both plans from 40 held-out states: the scorer's choice succeeds in 38, against 35 for the planner's own cost. On CAST, the scorer's choice among four saved proposals has an 11% lower trajectory error (Appendices K.2 and K.3). Figure 4a shows when the allocator keeps one full search.

Question 3 and Weakness 2: Shared and task-specific components, and Figure 3

Our framework shares one decision rule across all tasks: generate candidates, score them with a model trained on real execution outcomes, and either execute the best candidate or spend additional computation when every score is low. Each task then specifies three interfaces to this rule: what a candidate is, what the scorer reads and predicts, and what form the additional computation takes. These interfaces differ because the benchmarks differ in their actions and quality measures. HWM counts whether the agent reaches the goal, so the scorer predicts success. CAST measures trajectory error, so the scorer predicts the error. The LIBERO policy outputs a chunk of actions at a time, so a candidate is how much of the chunk to execute before calling the policy again. Section 4.4 and Table 4 (Appendix A.2, page 20) list the interfaces:

TaskCandidateScorer inputExtra computation when every score is low
HWM maze navigationplan from one full or two half-size searchesobservation, goal, predicted trajectory, computation usedanother world-model search
CAST visual navigationproposed pathcamera history, goal, predictionsmore proposals
LIBERO-Long manipulation10 or 16 actions of an action chunkimages and actionsan earlier policy call
ROOM navigation from imagesnavigation actionvisual history, goal, action summariescomparing the actions again with cached image features
A-EQA question answeringanswer now or after exploringquestion and viewsmore exploration
SAT spatial reasoninganswer with the reasoningimages, question, choices, question categorymore reasoning

Table 4 also lists the two supplementary video benchmarks.

The Figure 3 caption now states that one extra search runs when both plans score below 0.5, and explains that "Allocator loss" is Eq. 7 applied to the labels of the allocator. The allocator learns whether one full search succeeds and whether either of the two half-size searches succeeds; the scorer learns whether each plan succeeds (Section 4.6).

Best,
The Authors

Dear Reviewer ftBs,

Thank you for your precise questions, which helped us sharpen the paper considerably. With the discussion deadline approaching on November 18, tomorrow, we wanted to kindly follow up on our responses.

In brief, we showed that plans chosen when both candidates score low fail three times as often in re-runs, attributed the gains to allocation, scoring, and additional search on the same tasks and seeds, and made the shared decision rule and the task-specific interfaces explicit in Section 4.4 and Table 4.

Please do not hesitate to let us know if any point needs further clarification; we will reply as quickly as we can.

Best regards,
The Authors

Reviewer_R6sGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes an adaptive planning framework that combines outcome-grounded candidate comparison with selective computation allocation in world-model planning.

Critical strengths

The proposed approach evaluates candidate actions alongside their predicted consequences and determines whether additional search is necessary. The framework consists of a learned allocator that chooses between full-budget and split-budget searches, an outcome-supervised scorer that compares candidate plans, and a bounded additional-search mechanism triggered when the available candidates have low predicted success probabilities. The reported results show improvements in planning quality and computational efficiency.

Critical weaknesses
  1. The methodological novelty is not sufficiently established, particularly in relation to adaptive computation and existing sampling-based planning methods. the authors should clarify how their approach advances beyond existing adaptive computation strategies that allocate additional sampling or reasoning based on expected decision improvement.
  2. The paper lacks comparisons with simpler alternatives that could achieve similar improvements. A particularly important missing baseline is a standard MPPI planner equipped with a simple outcome predictor or success classifier.
  3. The outcome-grounded supervision requires substantial offline execution data, and its practical cost is not adequately considered.
  4. Several supplementary comparisons have limitations that weaken their relevance to the central claims.

Some minor issues: 1. The paper frequently describes the predicted success scores as evidence supporting execution. They are not the same. 2. Eq.7 combines several losses without sufficiently explaining their respective contributions in the main text. 3. The reference list contains inconsistent capitalisation of titles, conference names, and technical terminology.

Questions for authors

Please see the weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to R6sG

Dear Reviewer R6sG,

Thank you for your positive assessment and concrete suggestions; we are glad that you found no substantive concerns about soundness.

Weakness 1: Novelty relative to adaptive computation and sampling-based planning

StopTheDream differs from adaptive computation in what makes the decision: a scorer trained on real executions checks the plans the search has already finished, and the same score picks the plan and decides whether to search again. Section 2 now states the difference (page 3, lines 133–140):

ApproachThe decision is based onThe decision controls
Adaptive computation for planning [1, 2, 3]the agent's own estimate of the gain from more computationhow much to sample or predict
Compute allocation for language models [4, 5]the predicted difficulty or gain of each inputthe budget of each input, set before generation
V-GPS [6]a value function learned offlinewhich proposed action to execute
StopTheDreama scorer of finished plans, trained on real executionswhich plan to execute and whether to search once more

Deciding from real executions is more effective than deciding from the agent's own estimate, which existing adaptive computation uses. The agent's own estimate is least reliable for the plans that a cost-minimizing planner prefers (the optimizer's curse [7]). On CAST, we let the predictor's own cost, the predicted distance to the goal, make both decisions with the same computation as the scorer (Appendix K.3):

Decision maker on CAST (1,000 trajectories)Error at 17.5 predictions per trajectoryat 21.0at 25.5
Predictor's own cost1.5791.5851.577
Scorer trained on real executions1.4511.4311.412

Weakness 2: A simpler alternative, standard MPPI with a success classifier

We added the suggested baseline, and StopTheDream solves more tasks with about one seventh of the computation. The classifier is trained on the same 300 executions and added to the MPPI cost; the world model, tasks, and seeds are shared. One latent transition is one predicted step of one rollout (Section 5.2, Appendix K.2):

HWM, 120 tasksTasks solvedLatent transitions
MPPI with a success classifier in the cost1132,790.0 million
StopTheDream117384.4 million

The classifier costs more computation because, inside MPPI, every sampled plan must be decoded into low-level actions before the classifier can score the plan. The optimizer can also exploit errors of the classifier, as the optimizer exploits errors of the planner's own cost.

Weakness 3: Practical cost of the offline execution data

Collecting the 300 training executions and the 40 calibration states takes 18.7 hours of simulation once, and training takes about ten minutes on one GPU. The executions use 25 training maps disjoint from the test maps, need no human labels, and serve every test task (Section 4.4, Appendix K.1, page 39).

Weakness 4: Supplementary comparisons with limited relevance

We agree that the planning experiments carry the central claim; we added the supplementary benchmarks to show that the same decision rule also applies outside planning. Section 5.3 now states the role of the supplementary benchmarks (page 9, lines 470–472) and points to Appendix A.5 (page 20) for the protocols.

Minor points

  1. The revision calls the score "the predicted probability that executing a candidate succeeds"; "evidence" now refers only to the inputs of the scorer.
  2. Section 4.6 now states the role of each term of Eq. 7, and Appendix B.5 lists the auxiliary targets with the weight of each target.
  3. The reference list now uses consistent capitalization.

We hope the answers above resolve your concerns.

Best,
The Authors

References

[1] F. Callaway et al. Learning to select computations. UAI, 2018.

[2] S. Yu et al. When and how much to imagine: Adaptive test-time scaling with world models for visual spatial reasoning. arXiv:2602.08236, 2026.

[3] A. Z. Li et al. ELASTIC: Efficiently learning to adaptively scale test-time compute for generative control policies. arXiv:2606.31132, 2026.

[4] C. Snell et al. Scaling LLM test-time compute optimally can be more effective than scaling model parameters. ICLR, 2025.

[5] M. Damani et al. Learning how hard to think: Input-adaptive allocation of language model computation. ICLR, 2025.

[6] M. Nakamoto et al. Steering your generalists: Improving robotic foundation models via value guidance. CoRL, 2024.

[7] J. E. Smith and R. L. Winkler. The optimizer's curse: Skepticism and postdecision surprise in decision analysis. Management Science, 2006.

Dear Reviewer R6sG,

We are grateful for your positive assessment and constructive suggestions. Since the author–reviewer discussion closes tomorrow (November 18), we would like to follow up briefly.

Following your comments, we added the MPPI baseline with a success classifier, compared the outcome-trained scorer with the predictor's own cost on CAST, reported the one-time cost of the offline data, and clarified the role of the supplementary benchmarks.

We hope these additions resolve your concerns, and we are happy to answer any remaining questions before the deadline.

Best regards,
The Authors

Reviewer_rk16OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an adaptive world-model planning framework that dynamically adjusts search effort while evaluating potential plans, committing early when a viable plan is found and allocating more search only when initial rollouts are unpromising. The approach is tested on three benchmarks and shows some quantitative improvement over baselines.

Critical strengths

The high level pitch of the paper (at the level of the abstract, and beginning of the intro) is interesting: it frames a general problem for model-based simulation-driven planners, and presents an intuitively compelling approach. The experiments quantitatively support the hypothesis that this computation-sensitive, adaptive planning method can improve on baselines.

Critical weaknesses

I am not an expert in this topic, but I found the rest of the introduction and framing confusing in an important way: in deciding whether it is better or necessary to run more computation for search, is the question whether to search for an additional trajectory (because the current ones aren't good enough, as Figure 1 suggests, or is the question whether to run longer longer horizon rollouts (as Figure 2 suggests), or is the question whether to more generically search more (broadly or deeply) in a tree of possible action sequences? The framing is ambiguous.

I thought this would be resolved quickly when I went to read the technical sections of the paper, but they are very dense (including with a lot of notation that is not explained) and don't explain the approach clearly enough for me as a non-expert to understand, even well enough to resolve my confusion as described in the above paragraph. The results are only quantitatively summarized in tables; no specific examples are illustrated to provide insight or clarity on how the adaptive search works, which would also help readers understand better how and why the method works.

The results, while quantitatively better than baselines, don't seem to represent a dramatic or qualitative advance. Without a better understanding of how the method works, it's hard to judge how significant they are.

Also, the authors say that they used AI for all steps of the research. I am somewhat concerned that the weaknesses identified above are in part due to the use of AI. This makes me concerned about whether the authors actually understand their own system. At a minimum the authors should convince readers that they understand how their system works and what it is doing, but much better would be to revise the text so non experts can understand critical and basic issues such as where exactly the adaptive use of computation is focusing: longer roll outs, more roll outs, more possible plans, or some combination of these?

Questions for authors

If the authors can respond to any of the issues raised above in critical weaknesses, I am open to raising my score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer rk16 (1/2)

Dear Reviewer rk16,

We sincerely thank you for your time and for your candid and thoughtful review. We are glad that you found the high-level pitch interesting and the approach intuitively compelling, and we appreciate your detailed comments, which helped us improve the clarity of the paper. We address each of your concerns below.

Weakness 1: What the adaptive computation searches for (another trajectory, longer rollouts, or a tree)

Thank you for this question. StopTheDream decides whether to search for an additional trajectory when the current ones are not good enough, as Figure 1 suggests. It does not make the rollouts longer, and it does not search a tree. Each additional trajectory is one more plan for the agent to choose from; finding it takes more rollouts, and every rollout keeps the same length. We have clarified this in the introduction and at the start of Section 4 (page 4, lines 183–188), and the new Figure 4 shows one such decision on a real maze state.

Here is what these words mean in our paper. To make a plan, the planner tries out many action sequences in imagination: for each one, the world model predicts what would happen over a fixed number of steps. One such prediction is a rollout. The planner then keeps the action sequence that looks best. This is one plan, and the whole procedure is one search.

StopTheDream makes one decision at the start of each maze task, in four steps:

  1. Make one plan or two. The planner either runs one full search and gets one plan, or runs two searches of half the size and gets two plans. Both options cost the same. A small network (the allocator) picks between them (Section 4.6, Eq. 5).
  2. Check each plan. A second small network (the scorer) estimates, for each plan, the chance that it will reach the goal when executed. It learned this from 300 real runs that either succeeded or failed (Eqs. 6 and 7).
  3. Make one more plan, only if needed. If both plans have a chance below 0.5, the planner runs one more full search and gets a third plan. This happens at most once per task.
  4. Act, then continue with less computation. The agent executes the first four actions of the plan with the highest chance. From then on, the usual planner takes over and uses half as many rollouts at each step.

The first few actions decide which way the agent goes, and the later steps only follow that route, so they need fewer rollouts. Planners that use half as many rollouts at every step, without the decision at the start, solve 106 and 104 tasks. StopTheDream solves 117 with about the same computation (Section 5.2).

We spend the extra computation on more plans because the other two options help less. Using twice as many rollouts at every step costs twice the computation and solves fewer tasks (Figure 4d). Making the rollouts longer stops helping beyond a certain length (Figure 2), so we keep the length fixed.

Weakness 2: Dense technical sections and unexplained notation

Thank you for pointing this out. We have revised Section 4 so that it first describes the decision in plain language before introducing any notation (page 4, lines 188–194). We have also added a notation table (Table 3, Appendix A.1) that explains the meaning of every symbol used in the paper, so that readers can look up any symbol without searching the text.

Weakness 3: A concrete example of how the adaptive search works

Thank you for this suggestion. We have added a new Figure 4 (page 5) that follows one real decision on a held-out HWM maze state, step by step.

  1. Two plans, both unlikely to succeed. Two half-size searches give two plans. The scorer estimates their chances of reaching the goal at 0.088 and 0.006, both below the 0.5 needed to execute a plan.
  2. One more search. StopTheDream therefore runs one more full search. The new plan gets a chance of 0.9995, far above the earlier best plan (0.079 when the two are compared).
  3. Execution. The agent executes the first four actions of the new plan, and the planner then continues with half its usual rollouts. The executed route reaches the goal in 2 of 3 runs.
AuthorsOpenReview ↗

Response to Reviewer rk16 (2/2)

Weakness 4: Significance of the improvements

Thank you for raising this point. The main improvement is in computation: on the two planning benchmarks, StopTheDream needs about half the computation of the strongest baseline and still reaches better quality. On the three other benchmarks, it keeps or slightly improves the quality of the strongest baseline with 15% to 39% less computation. We summarize the results below (Table 1, Figure 5, Appendices E, G, and H). Each benchmark counts computation in its own unit; on HWM, one latent transition is one predicted step of one rollout.

Benchmark (strongest baseline)Quality: baseline vs. StopTheDreamComputation: baseline vs. StopTheDream
HWM maze navigation, 120 tasks (CoVO-MPC)113 vs. 117 tasks solved747.5 vs. 384.4 million latent transitions; 25,037 vs. 18,934 s
CAST visual navigation, 1,000 trajectories (PiJEPA)trajectory error 1.4224 vs. 1.362928.0 vs. 12.85 predictor calls per trajectory
LIBERO-Long robot manipulation, 1,000 episodes (chunk length fixed per task from training data)92.40% vs. 92.83% success22.25 vs. 20.85 policy calls per episode
LIBERO, four suites, 2,000 episodes, schedule calibrated per suite (FasterWAM with feedback every 10 actions)98.75% vs. 99.05% success31,438 vs. 23,445 policy calls
ROOM navigation from images, 100 start states (DreamerV3)100 vs. 100 goals reached36.0 vs. 22.1 s of policy time
A-EQA embodied question answering, 184 questions (explore before every question)answer score 48.913 vs. 48.913184 vs. 156 explorations; 31,626 vs. 26,120 s

We also compared at about the same computation. When CoVO-MPC, the strongest HWM baseline, runs with half its budget, it uses 376.8 million latent transitions and solves 106 tasks, against 117 for StopTheDream with 384.4 million (Section 5.2).

Weakness 5: Use of AI and understanding of the system

Thank you for raising this concern directly. The idea behind StopTheDream is ours, and it comes from one observation that the paper tests: the planner's own cost is least reliable for exactly the plans the planner prefers. The planner's own cost is the predicted distance between the final state of a plan and the goal. The planner searches for the plan with the lowest cost, so the plan it picks is often one where the world model is wrong. This is why StopTheDream checks finished plans with a scorer that has learned from real runs, and searches once more only when every plan looks unlikely to succeed. Our answer to Weakness 1 describes the full procedure step by step.

Each design choice in StopTheDream is backed by a test of what happens without it:

  • The scorer must read the actions of a plan. When the scorer reads only the imagined trajectory, its picks are as bad as random picks on CAST (trajectory error 1.619 against 1.620). When it also reads the actions, the error drops to 1.532 (Appendix K.3).
  • The extra search should run only when both plans score low. When it runs on all 99 tasks with two plans, StopTheDream solves 115 tasks. When it runs only on the 11 tasks where both plans score below 0.5, StopTheDream solves 117 (Table 2).
  • The scorer should check finished plans, outside the planner's search. When a success classifier is placed inside the search, the planner needs 7.3 times the computation and solves 113 tasks, against 117 for StopTheDream (Appendix K.2).

As our AI use statement says, AI tools assisted with code writing, discussion of the mathematical formulation, feedback on experimental designs, interpretation and presentation of our results, literature discovery, and drafting and revising the manuscript. We decided what to build and which experiments to run, and we are responsible for the final text, claims, code, measurements, and citations. We agree that the original presentation made the method hard to follow, and we have revised Section 4 and added Figure 4 so that readers can see what the system does.

We hope the answers above resolve your concerns.

Best,
The Authors

Dear Reviewer rk16,

Thank you again for your candid review. As the discussion period ends tomorrow, November 18, we wanted to kindly check whether our responses have addressed your concerns.

In particular, we clarified that the adaptive computation is spent on additional candidate plans, with a fixed rollout length and no tree search. We also rewrote Section 4 in plain language before any notation and added a real step-by-step example in the new Figure 4.

If anything remains unclear, please let us know, and we will respond promptly to any further questions.

Best regards,
The Authors

Dear Reviewer rk16,

We would like to kindly follow up on our response, as the discussion period ends soon (AOE end of day, November 18th).

If anything is unclear, we are happy to provide further clarification and responses, and we will respond promptly to any further questions.

Thank you so much.

Best,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and All Reviewers

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all reviewers for the careful and constructive reviews. We appreciate the positive feedback on our motivation and the intuitive appeal of our approach (rk16), the practical relevance of the problem (ftBs), and the soundness of our method (R6sG). In response, we have revised the manuscript to improve its clarity and added new analyses. All changes are highlighted in red in the revised PDF, and the results reported in the original submission remain unchanged. Below, we summarize our responses to the main concerns.

Clarity of the method

  • Where the extra computation goes (rk16, ftBs). In our planning experiments (HWM and CAST), the extra computation is used to search for more candidate plans. It does not extend the prediction horizon, and there is no tree search. We have clarified this in the introduction and at the start of Section 4, and the new Figure 4 follows one real HWM decision in which two low-scoring plans trigger one more search.
  • Readability of Section 4 (rk16). Section 4 now begins with a plain-language description of the method before any notation, and we have added a notation table (Table 3, Appendix A.1) that explains every symbol.
  • Shared and task-specific components (ftBs). Section 4.4 and Table 4 now separate the decision rule shared by all tasks from the three parts that each task specifies: the candidates, the scorer's input and target, and the form of the extra computation. The caption of Figure 3 now states the condition for the additional search and the training targets of the allocator.

New experiments and analyses

  • Simpler baseline (R6sG). We added a standard MPPI planner with a success classifier in its cost. It solves 113 HWM tasks, against 117 for StopTheDream, and uses 7.3 times as many latent transitions (Section 5.2, Appendix K.2).
  • Novelty relative to adaptive computation (R6sG). StopTheDream uses a scorer trained on real executions both to select a plan and to decide whether to search again; Section 2 now states how this differs from prior work. On CAST, the scorer gives a lower trajectory error than the predictor's own cost at the same number of predictions (Appendix K.3).
  • Evidence for searching again on low scores (ftBs). On HWM, plans chosen when both candidates score low fail three times as often in re-runs (18% against 6%). Searching again only in these 11 tasks solves 117 tasks, against 115 when searching again in all 99 tasks with two plans. A task-by-task comparison with the same seeds shows that the additional search rescued 1 task and lost none. All 4 failures without the additional search fell in the 11 low-score tasks. (Section 5.5, Appendix K.2).
  • Contribution of each component (ftBs). The additional search contributes one solved task (116 to 117) for 0.30% more computation. Allocation and scoring contribute the rest of the success gain (116 against 105 for the original HWM planner). The computation saving comes from replanning with half the usual samples after the first four actions (Tables 1 and 2, Appendix K.2).
  • Cost of the offline data (R6sG). On HWM, collecting the training and calibration data takes 18.7 hours of simulation once, and training the scorer and the allocator takes about ten minutes on one GPU (Section 4.4, Appendix K.1).

Significance and scope

  • Significance of the improvements (rk16). On the two planning benchmarks, StopTheDream reaches better quality with about half the computation of the strongest baseline. On the other benchmarks in Figure 5, it matches or exceeds the strongest baseline with less computation. At about the same computation, the strongest HWM baseline solves 106 tasks, against 117 (Section 5.2).
  • Role of the supplementary benchmarks (R6sG). Section 5.3 now states that the planning experiments test our central claim, and that the supplementary benchmarks show the same decision rule applied outside planning.

We respond to each reviewer in detail below.

Best,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.