acceptodds
Under review as a conference paper at ICLR 2027

Which Predictions Suffice for a World-Model Planner's Decision? Certifying Reward and Value Evidence by Re-Planning

Abstract

World-model planners allocate search or optimization among actions using predicted rewards and values. We study the minimum prediction evidence sufficient to preserve a decision's action and preference statistics under native re-execution. Two challenges arise: (1) Frozen computation. Scoring recorded computation omits the planner's adaptation to replaced predictions. (2) Single-set explanation. One passing set leaves minimum size, alternatives, shared evidence, and reliability across planning streams unresolved. We introduce , a sufficiency-certification framework whose interventions govern recorded and newly visited prediction read sites. Exact auditing recovers the minimum size, all minimum certificates, and their core; budgeted search and constrained repair return verified sets. Certificates are execution-conditioned, while independent calibration certifies cross-stream reliability with a simultaneous finite-sample guarantee. We prove that the original transcript alone cannot generally decide sufficiency and that execution-conditioned and stream-reliable minimum sizes can differ by \Theta(H). Across four planners and six tasks, native re-execution rejects 78.9% of the tested frozen-candidate explanations on TD-MPC2 and 16.5% of frozen-tree explanations. At a 384-query cap, default FutureCert recovers the exact minimum size on 93.0% of TD-MPC2 decisions, compared with 86.5% for its Beam-4 variant. Exact audits expose alternative support and repair opportunities; cross-stream analysis measures the additional evidence required for reliability.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, Reinforcement Learning, Planning, Decision Certificates, Monte Carlo Tree Search

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wGPcOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies which predicted rewards and values suffice to preserve a world-model planner's action and preference statistics. It introduces FutureCert, which retains selected prediction-read groups, replaces the rest with fixed references, and reruns the native planner, including newly visited prediction sites. The framework supports exact minimum-certificate auditing, budgeted search and repair, and independently calibrated reliability across planning streams. The main evaluation covers four planners and six tasks.

Critical strengths
  • The problem is very much well motivated, and the experiments are carefully designed. Changing reward or value predictions can change what the planner explores next, so an explanation evaluated on frozen computation might not survive actual re-planning. The paper supports this argument with both a concrete counterexample and empirical evidences. Importantly, the authors distinguish a limitation of the frozen interface from a limitation of the attribution methods themselves. This makes the evaluation substantially more informative.
  • The different guarantees are carefully separated. Finding a sufficient set, proving minimum size, recovering all minimum alternatives, and establishing reliability across streams are treated as different questions.
Critical weaknesses
  • The explanatory value of the richer certificates is less established than their formal validity. The certificates seems like based on prediction reads under a specific intervention instead of particular imagined futures. The sensitivity analysis is useful, but it remains unclear which audit choices yield the most meaningful explanation for a practitioner. I would view this as a limitation of the demonstrated usefulness, rather than a flaw in the stated guarantees.
  • Scalability might remain a meaningful limitation. Exact family/core audits mainly cover H <= 12, with 4,096 verifier queries for the complete enumeration at H=12, plus additional calibration runs. The large-space experiments are useful, but do not provide the same complete exact ground truth. The default search also trades additional preparation cost for better recovery like what the paper transparently reports.
Questions for authors

Q1: Beyond the action/ranking/margin controls, can you separate frozen-tree failures involving substantial changes among leading actions from those caused only by reordering low-visit actions?

Q2: For the larger evidence space, how does verified certificate size trade off against total cost?

6 comments
AuthorsOpenReview ↗

Reply to Reviewer wGPc

Thank you for your assessment of the motivation and experimental design. We agree that the value of a certificate depends on the decisions a practitioner can make with it. Tables R1–R18 are the appendix tables added during review.

AuthorsOpenReview ↗

W1 Explanatory value of the richer certificates

If a group of reads becomes unavailable, can the planner preserve this decision with the same number of other groups? For MuZero Pong decision 81, the two minimum supports are {1, 4, 6} and {3, 4, 5}, each of size 3. If unit 1 is unavailable, the second support still works. If unit 4 is unavailable, all 26 candidates of size at most 3 that avoid it fail. The smallest feasible support now has four units, {1, 3, 5, 6}.

The certificate identifies groups of prediction reads for a specified input, planning stream and replacement rule. On re-execution, retained groups use the predictions actually produced, and the remaining reads receive their reference values. The resulting action and preference statistics must pass the fidelity check. The same group also covers reads at newly visited sites, so the same support can admit different imagined trajectories. The Core consists of the groups shared by every minimum support.

Across four TD-MPC2 tasks, the 692 inputs with nonempty minimum supports yielded 8,304 single-unit loss scenarios. Of the 6,780 scenarios admitting a same-size support, the first certificate covered 6,121. The complete family covered another 659. For the remaining 1,524/8,304 scenarios (18.4%), the complete family ruled out every same-size substitute. Any feasible alternative would have to be larger.

These queries refer to the same input and audit specification. Once the family has been stored, a practitioner can select a verified substitute without another verifier call (Table R15).

Which tasks warrant a closed-loop compression test? Across five tasks, both leave-one-episode-out predictions on the development set and closed-loop measurements ranked Finger Spin first. The predicted saving was 70.3% of prediction reads, and the measured saving was 50.9%. On Cheetah Run, Cup Catch, Walker Walk and Quadruped Run, both figures were at most 4.1%.

The choice also depends on return. On Finger Spin, skipping reads according to the certificate took 0.77 times the full planning time and reduced return by 34.6. Reducing planning iterations from 6 to 4 took 0.79 times the full planning time and reduced return by 1.9. At these comparable runtimes, reducing the iteration count preserved more return.

In an exploratory DreamerV3-S training experiment, we assigned larger weights in the critic’s TD loss to imagination steps marked by a certificate. We compared equal weights and a random permutation of the same certificate weights. Each run used 100k environment steps. We averaged return over each training run, then averaged over three seeds per task.

Table WG-R1. Mean return over training

TaskCertificate weightingEqual weightingShuffled weights
Cup Catch844775789
Cheetah Run333290315
Hopper Stand254178199

Certificate weighting gave higher mean returns than either control on all three tasks (Table R6). With three seeds per task, we treat this training comparison as exploratory. The reconfiguration experiment above directly measures what the complete family adds over one certificate.

Practitioners can choose the fidelity criterion to match the quantity they need to preserve. They can constrain action deviation alone, or also constrain the first step of the optimized action mean. Across 800 inputs in four tasks, mean minimum support size under the joint criterion was 4.19, 2.73 and 1.92 units at tolerances ε = 0.1, 0.2 and 0.3, respectively (Table R12).

The groups should match the reads that can actually be disabled. For comparisons across partitions, the retained groups can be expressed in the original units. A joint support can be computed when several specified reference values must all be accommodated. Whether these choices improve practitioners’ understanding or diagnostic performance remains an open empirical question.

In the revision, §3 and Fig. 1(b) will illustrate the read groups and their alternatives. Section 5.4 gives the gains from storing alternatives. Section 5.7 brings the five applications together in Table 4 and compares predicted and observed closed-loop read savings in Table 5. Supplementary Tables R6, R12, and R15 give the corresponding comparisons.

AuthorsOpenReview ↗

W2 Scalability

The 24-unit native audit recovered the complete minimum family and its Core on 136/160 inputs (85.0%). On the other 24 inputs, every candidate of size 0–5 failed, establishing the lower bound k⋆≥6k^\star\ge6. Completion ranged from 25/40 on Quadruped Run to 40/40 on Finger Spin.

We took 40 inputs at fixed indices in each of four TD-MPC2 tasks and refined the original partition into 24 units. The audit reruns native MPPI for each mask. It checks sets in increasing order of size and finishes the first layer containing a passing set, up to layer 5. On inputs it completes, this layer-by-layer audit uses ∑j=0k⋆(nj)\sum_{j=0}^{k^\star}\binom{n}{j} mask checks, which is O(nk⋆)O(n^{k^\star}) for fixed minimum size and 2n2^n in the worst case.

Across all 160 inputs, the audit averaged 18,382 checks per input (median 2,325), including 55,455 checks for each unfinished input. On shared H200 GPUs with six processes per GPU, the consistency checks and both the 12- and 24-unit exact audits together took a mean of 1,143 seconds per input (median 166 seconds), excluding model loading (Table R13).

Budgeted search returned natively verified sets on the same 160 inputs. Table WG-R2 reports recovery and Gap on the 136 inputs with known minima. Raising the cap from 384 to 768 reduced mean excess size on these inputs from 1.60 to 0.79 units. Q2 gives the sizes returned on the remaining 24 inputs.

Table WG-R2. Budgeted recovery on inputs with known minimum size

Query capInputs recovering minimum sizeMean excess units (Gap)
38464.7%1.60
76877.2%0.79

A separate 12-unit experiment measures the cost of preparation. On the 600 TD-MPC2 inputs in the paper’s Table 2, Beam-4 requires no preparation. The default configuration reuses a development library for each task and recovers minimum size more often.

Table WG-R3. Preparation cost in the 12-unit experiment

ConfigurationMinimum-size recoveryGapMean online queries/inputMean preparation queries/task
Beam-486.50%0.30769.440
Default92.97%0.14069.6217,047.3

Table 2 weights episodes and settings equally and averages the default configuration over ten solver seeds. If we audit NN inputs in each of the three tasks and reuse the development library within each task, the default costs an average of 69.62+17,047.3/N69.62+17{,}047.3/N queries per input. This gives 154.86 at N=200N=200 and 86.67 at N=1,000N=1{,}000.

Certifying reliability on fresh planning streams incurs an additional calibration cost. On the original 20 Finger Spin inputs, the broader candidate pool averaged 757.5 non-full calibration executions per input, plus 64 shared full-retention executions and the cost of proposing candidates.

Section 5.3 will report the 24-unit completion rate. The 12-unit main table will give online and amortized costs side by side. Supplementary Tables R7 and R13 distinguish search preparation, exact auditing and calibration costs.

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

This paper addresses the problem of explaining decision-making in world-model-based planners. It proposes FutureCert, a framework for verifying the sufficiency of reward and value prediction evidence, identifying minimum sufficient certificates, analyzing the complete family of minimum certificates and their common core, and searching for and repairing certificates under a limited query budget. The experiments cover several planning configurations, including MuZero, Sampled EfficientZero, EfficientZero, and TD-MPC2. The results show that fixed-trace evaluation can overestimate evidence sufficiency, while minimum-certificate analysis identifies alternative sufficient evidence sets and their shared core.

Critical strengths
  1. The paper addresses a limitation of analyses based on fixed search trees or candidate trajectories. These approaches may fail to capture how a planner's search or optimization changes after predictions are replaced. By rerunning the native planner and continuing to intervene on predictions at newly visited sites, FutureCert provides a more faithful assessment of certificate sufficiency under the specified intervention setting.
  2. The paper extends certificate sufficiency verification to the analysis of minimum-certificate structures, including the minimum certificate size, the complete family of minimum certificates, and their common core. It also considers certificate repair when evidence is unavailable. This allows the framework to identify alternative sufficient evidence sets and examine which evidence is shared across all minimum certificates.
  3. The experimental evaluation covers a range of planning configurations and tasks. The paper reports nine experimental settings involving MuZero, Sampled EfficientZero, EfficientZero, and TD-MPC2, spanning tree-search-based planning and continuous control tasks.
Critical weaknesses
  1. The conclusions drawn from certificate sufficiency analysis depend on the specified audit setting. FutureCert replaces selected predictions and reruns the planner, but the resulting certificates may be affected by how evidence units are partitioned, which reference values are used for replacement, and how sufficiency is defined. Although the paper examines sensitivity to some of these choices, it remains unclear how stable the certificate structures and their interpretations are across different reasonable audit settings.
  2. Exact audits are currently limited to small evidence spaces, as exhaustive enumeration of the complete family of minimum certificates becomes increasingly expensive with more evidence units. The scalability of FutureCert for larger planning problems therefore remains unclear, particularly for recovering the complete certificate family.
Questions for authors
  1. How is the scalability when the number of evidence units exceeds 12? Is there a corresponding approximate guarantee or layered audit strategy?
  2. Beyond identifying a single minimum certificate, the paper analyzes the complete family of minimum certificates. Does access to the complete family provide measurable benefits for diagnosis, certificate repair, or query efficiency compared with using only one minimum certificate? When the query budget is insufficient to enumerate the complete family, to what extent can a partial set of minimum certificates retain these benefits?
  3. The paper uses a fixed reference value to replace removed predictions. What is the rationale for this choice, and in what sense does it better capture the intended notion of evidence removal than zero replacement, random replacement, replacement based on a conditional distribution, or adversarial replacement? Could the authors provide a theoretical justification or empirical comparisons showing how these alternatives affect certificate size and structure?
8 comments
AuthorsOpenReview ↗

Reply to Reviewer sqUj

Thank you for raising these questions about the audit setting and scalability. We agree that interpreting and using a certificate requires knowing how sensitive it is to the choices made in the audit. Tables R1–R18 are the appendix tables added during review.

AuthorsOpenReview ↗

W1 Dependence on the audit setting

Changing the reference, grouping or tolerance can change both how many reads suffice and which reads are needed. Reference changes often preserved individual certificates on Cheetah Run. On Cup Catch, requiring one support to pass under both references increased the minimum size in 16 of 30 input–draw pairs.

1. Reference values. We resampled the development episodes used to estimate the per-unit mean reference three times. For the same 10 inputs per task, we recomputed the minimum family under each new reference and tested whether an original minimum support still passed. The resulting 30 pairs per task share inputs.

Table SQ-R1. Native FP32 MPPI under resampled references

TaskMean k⋆, original / newPairs with equal k⋆Original minimum certificates passing under new referenceCore Jaccard (eligible pairs)Pairs retaining original k⋆ under both references
Finger Spin0.70 / 0.7030/3080/84 (95.2%)0.667 (6)30/30
Cheetah Run2.80 / 2.8329/3069/75 (92.0%)0.898 (30)29/30
Cup Catch3.10 / 3.1722/3021/52 (40.4%)0.400 (26)11/30 confirmed

Transfer pools the original certificates evaluated under each draw. On Cup Catch, the joint minimum equaled the original in 11 pairs and increased in 16. In the other 3, evaluation labels for the original certificates under the new reference were unavailable. The 21/52 transfer rate uses the 27 evaluated pairs. Core Jaccard includes pairs with at least one nonempty Core. On Finger Spin, both Cores were empty in 24 pairs, including 15 where the empty support passed both references.

2. Evidence partition. In four TD-MPC2 tasks (200 inputs each), we evaluated three ways to group the original 12 units, holding the reference and fidelity criterion fixed. G6 groups reward versus terminal value by three iteration stages, Q4 groups by predicted quantity, and P3 groups by iteration stage. We minimized the number of groups in each partition and expressed their size in original units.

Table SQ-R2. Partition sensitivity on the same 800 inputs, averaged equally over tasks

PartitionMean original units retained (of 12)Inputs with a unique minimumCoarse minima containing a fine minimumFine minima whose projection is a coarse minimum
12 units (identity)2.7354.2%100%100%
G65.6474.8%76.5%41.8%
Q46.9685.5%84.7%39.1%
P38.9390.2%96.5%67.4%

Grouping can merge distinct alternatives. For example, grouping by iteration stage raised the fraction of inputs with a unique minimum from 54.2% to 90.2%, while increasing the mean number of original units retained from 2.73 to 8.93. W2 reports refinement from 12 to 24 units.

3. Fidelity tolerance. On the same 800 inputs, we re-evaluated all 4,096 masks. By default, both the executed action and the first step of the planning mean, μ₀, must stay within ε = 0.2 of their full-evidence values in L∞ norm.

Table SQ-R3. Fidelity criteria relative to ε = 0.2, averaged equally over tasks

CriterionMean k⋆ (of 12)Inputs with unchanged k⋆Mean absolute change in k⋆Core Jaccard
ε = 0.14.1929.2%1.460.363
ε = 0.2 (reference)2.73100%0.001.000
ε = 0.31.9246.6%0.810.325
Action deviation ≤ 0.2 only2.3070.1%0.430.540

Tightening ε to 0.1 required 1.46 more units on average. Relaxing it to 0.3 saved 0.81 units. Relaxation preserves previously passing sets, but the minimum supports and their shared members can change. Core Jaccard includes inputs with at least one nonempty Core.

Section 3 will specify the three audit choices. Section 5.6 will distinguish changes in minimum size from changes in membership. Tables R10–R12 give the sensitivity analyses and Q3 compares additional replacement rules.

AuthorsOpenReview ↗

W2 Scalability beyond small evidence spaces

After refining the evidence partition from 12 to 24 units, native MPPI auditing established the exact minimum size, complete minimum family and Core on 136/160 inputs (85.0%). On the remaining 24 inputs, every mask of size 0–5 failed, proving that at least 6 units were necessary.

We took 40 inputs at fixed indices 0, 5, …, 195 from each of four TD-MPC2 tasks. For every mask, we reran MPPI so that the intervention could change later candidates. We enumerated masks by increasing size, stopping after completing the first layer containing a passing mask or reaching the cutoff at layer 5. The audit recovered exact results on 40/40 Finger Spin inputs, 37/40 Cheetah Run inputs, 34/40 Cup Catch inputs and 25/40 Quadruped Run inputs.

Across all 160 inputs, the audit used a mean of 18,382 native verifications per input and a median of 2,325. Each of the 24 unfinished inputs counts as 55,455 verifications. On shared H200 GPUs with six processes per GPU, the consistency checks and the exact 12- and 24-unit audits together took a mean of 1,143 seconds per input and a median of 166 seconds, excluding model loading (Table R13). Q1 gives the size lower bound and worst-case query count.

When one small verified support suffices, budgeted search can stop earlier. On the same 160 inputs, search reused each task’s 12-unit development library and prior. At caps of 384 and 768, the mean numbers of online queries charged to the respective caps were 224.3 and 356.9. The mean returned sizes were 5.16 and 4.44 units.

On the 136 inputs with known minima, search recovered the exact minimum size on 64.7% and 77.2% of inputs at the two caps, with mean excess sizes of 1.60 and 0.79 units. On the remaining 24 inputs, the mean returned sizes were 11.17 and 10.96 (Table R18).

A coarse certificate can also serve as the starting point for finer search. On 60 inputs each from Cheetah Run and Cup Catch, we mapped the default 12-unit certificates to the 24-unit partition and continued searching. Of these inputs, 20 overlap with the 160 above. The cap of 384 online queries covers obtaining the mapped starting point and refining it. Offline preparation of the development library is accounted for separately.

Search used 272.3 queries on average and reduced the mean retained fraction from 25.4% for the mapped supports to 20.4%. All returned supports passed native verification. For each solver seed, we computed the fraction of inputs with a strictly smaller support within each task and then averaged the two tasks equally. Across five seeds, the mean was 66.2% and the standard deviation was 0.7 percentage points.

In §5.3, we will report the exact completion rate and the lower bound on the harder 24-unit inputs, and update the limitations accordingly. Supplementary Tables R13 and R18 give the proof cost by task and the budget sweep on the same inputs.

Reviewer_UCt1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents FutureCert, a framework for identifying minimum sets of predicted rewards and values that are sufficient to preserve the decisions of world-model planners. The method evaluates sufficiency by re-executing the planner while retaining selected predictions and replacing others with fixed reference values, accounting for changes in the planning process. The framework considers both the selected action and the planner's preference statistics. It introduces an exact auditing procedure to identify minimum certificates and their shared evidence, a budgeted search method for larger spaces, and a calibration procedure to assess reliability across different planning settings. The approach is evaluated on six tasks using MuZero, Sampled EfficientZero, EfficientZero, and TD-MPC2. Additional experiments investigate the differences between frozen and native verification, minimum certificate recovery, and the robustness of the resulting certificates across planning conditions.

Critical strengths

• Interesting Problem: The paper addresses an interesting problem in explaining the decisions of world-model planners. Re-executing the planner allows the evaluation of prediction sufficiency under changes in the search process. The experiments show that this distinction can substantially affect the validity of explanations. • Relevance to the conference: The topic of the paper is relevant to the ICLR audience. • Technical soundness: I couldn't go in depth in all technical details but I did not find specific errors in the central formal claims. • Adequate experiments: The experimental evaluation covers different planning architectures and tasks, showing substantial differences between frozen and native verification. The analysis of minimum certificates and alternative solutions provides additional insights into the evidence supporting planning decisions. The independent calibration procedure also provides finite-sample reliability guarantees under the stated assumptions. • Reproducibility: The paper provides an anonymous code repository for review and gives substantial protocol detail, including evidence partitions, tolerances, seed splits, and evaluation populations. The declared plan to release the complete code and artifacts upon acceptance is helpful.

Critical weaknesses

• Novelty/originality: The main contribution is a useful integration and formalization of native planner re-execution, minimum sufficient-set structure, repair, and reliability calibration. However, many ingredients build on established ideas: sufficient subsets and backbones, perturbation-based attribution, native intervention, and binomial calibration. The paper should state more sharply what is novel beyond combining these ingredients for adaptive planners, and separate that contribution from the exact-enumeration and search procedures. This is the main weakness of this paper (together with significance/impact, see below), that makes me give a slightly negative score to the paper. The main contribution appears to be the integration of planner-native re-execution with minimum-certificate auditing and cross-stream calibration, but the paper does not sufficiently establish how much this advances beyond prior work on sufficient explanations, minimum explanation families, and calibrated sufficiency. The experiments demonstrate that the framework functions as an audit, yet the adapted baselines and substantial preparation costs make its advantage over existing approaches difficult to assess. The authors should sharpen the novelty claim and provide a clearer account of the practical value of this integration. See question 0, this point is important to address my final rating. • Significance/impact: The framework’s guarantees are tied to a specific audit specification and a fixed replacement rule; they do not establish sufficiency under all ways of completing removed predictions, nor human usefulness of the explanations. Practical benefit is also uneven. The reported fallback deployment saves substantial computation on Finger Spin but loses some return; Cheetah Run and Cup Catch mostly retain full planning. The paper should make clearer which users or applications benefit enough to justify the audit and calibration cost. • Writing quality and organization: Despite careful documentation, the paper is exceptionally long and dense. It introduces many definitions, solver variants, controls, and appendices, making it difficult to identify the central conceptual contribution and the most decision-relevant results. A shorter main narrative, a clearer separation of core findings from diagnostic studies, and , mainly, a compact running example, would improve readability. • The related work omits Alkhatib and Lowry’s recent study, which uses conformal prediction to turn attribution-selected feature subsets into sufficient explanations with confidence guarantees. While their method concerns image classifiers rather than adaptive planners, its close connection to this paper’s calibration of sufficient subsets should be acknowledged and the distinction made explicit. Reference: Amr Alkhatib and Stephanie Lowry. “Turning Feature Attributions into Sufficient Explanations Using Conformal Prediction.” Proceedings of the Fifteenth Symposium on Conformal and Probabilistic Prediction with Applications, Proceedings of Machine Learning Research, vol. 329, pp. 382–401, 2026. • Certificate sensitivity/stability/usefulness: The certificates are sensitive to the conditions under which they are computed. Table 3 shows substantial differences across planning streams and numerical implementations, with a minimum-family Jaccard similarity of 0.179 on Cup Catch when changing the arithmetic implementation. Moreover, the cross-stream evaluation certifies non-full sets on only 1/50 Cheetah Run inputs and 0/50 Cup Catch inputs. These results raise concerns about the stability and practical usefulness of the resulting explanations. • Computational cost: The computational cost remains a limitation, particularly for larger spaces. While FutureCert achieves higher minimum-size recovery than Beam-4 (92.97% versus 86.50%), its amortized cost is also higher (154.86 versus 69.44 queries). Furthermore, Section 5.4 shows that the first four minimum certificates already recover 2,043 of 2,045 substitution opportunities. This leaves the additional benefit of complete-family enumeration unclear relative to its computational cost.

Questions for authors
  1. Could you clarify which parts of FutureCert constitute the paper’s novel methodological contribution beyond prior work on sufficient explanations, MCTS explainability, and conformal calibration of sufficient subsets? In particular, which experiments isolate the benefit of planner-native re-execution at newly visited read sites from the benefits of the existing minimum-set and calibration components? 2. Table 3 shows that certificate structure can change substantially across planning streams and numerical implementations. Which properties of the certificates remain stable under these changes, and how does this affect their interpretation? 3. How does FutureCert scale with an increasing number of evidence units? In particular, what are the main computational limitations when moving towards more fine-grained explanations? 4. Could the authors provide a clearer comparison of computational costs, including the offline preparation required by FutureCert, against simpler approaches using native verification? 5. Section 5.4 shows that just a few minimum certificates recover almost all substitution opportunities. What is the practical benefit of computing the complete family rather than a small number of verified alternatives? 6. Please elaborate on the weaknesses.
16 comments
AuthorsOpenReview ↗

Reply to Reviewer UCt1

Thank you for recognizing the value of native re-execution. We agree that the paper should explain more clearly what this design adds to prior work and how a developer can use it. Tables R1–R18 are the appendix tables added during review.

AuthorsOpenReview ↗

W1 Novelty/originality

Thank you for pointing out that describing FutureCert as a “combination” left unclear what new problem arises in an adaptive planner. We define sufficiency over the entire planning execution after intervention. This lets us determine which reads suffice, which can substitute for one another, and whether a minimum set for one planning stream can be reused on fresh streams. Three findings distinguish this problem from earlier ones.

1. Sufficiency must cover sites first visited after intervention. Lam et al. (2022) use search trees and counterfactual states to test rules about planning inferences. We ask whether retaining specified reward and value predictions preserves the action and preferences when the planner searches again. Replacing a prediction can redirect search into a new branch, where reads of a removed group must also receive the reference value. Proposition 1 shows that two models can agree on the entire original record yet give different sufficiency verdicts after intervention. Of 4,518 non-full candidate sets across three MuZero games, our record check reached a verdict on only 109 (2.4%). It returned “unknown” on the rest, which required native re-execution for a verdict. Q1(b) tests the effect of replacement at newly visited sites in isolation.

2. Even when every single-group deletion fails, a smaller sufficient set may exist. Under the monotone predicate of Ignatiev et al. (2019), deleting one group at a time can yield an explanation with no sufficient proper subset. In a planner, every single deletion can fail even though removing several groups together redirects search and preserves the decision. On 600 TD-MPC2 inputs, 42.0% of the sets at which single-group deletion stopped still contained a proper subset that passed verification. We therefore check sets in increasing order of size to find all supports of minimum cardinality and the members they share (the Core). Families of minimal explanations and their shared members have been studied before (Marques-Silva, 2022). Here we measure the alternatives they expose for a planning decision. On 152 tree-search inputs with multiple minimum certificates, the first certificate covered 1,402 single- or double-group loss scenarios admitting a same-size support. The first four covered 2,043, and the complete family covered all 2,045.

3. The number of reads needed for one execution can be much smaller than the number needed for reuse. Alkhatib and Lowry (2026) calibrate the probability that an attribution-selected subset preserves the prediction on a new input. We fix the input and ask whether the same preselected reads preserve the decision on fresh planning streams. Proposition 2 constructs a planner in which the two minimum sizes differ by an amount that grows linearly with the number of read groups. For Cheetah Run input 83, solving each stream separately gave a median minimum of 4 groups. Requiring one set to pass on at least 13 of 16 discovery streams raised the exact minimum to 12 groups. Five of six diagnostic inputs required at least 10 groups under the latter criterion. We separately certify the probability of passing on future streams using independent samples. This certifies a preselected candidate’s reliability without establishing that it is the smallest among all reliable sets.

We will make these contributions explicit in the Introduction and Related Work. We use established tools for solving and certifying the resulting problems, including layer-by-layer enumeration, budgeted search, and binomial confidence bounds. Table R8 gives the questions addressed by four lines of prior work and the evidence for our contribution in each case.

AuthorsOpenReview ↗

W2 Significance / impact

Thank you for asking what a developer can do with a certificate. For an audited decision, the complete minimum family tells us which losses admit a same-size substitute and which rule out every support of that size. We also tested four uses of the selected imagination steps.

Table UC-R1. Applications and measured outcomes

ApplicationEvaluationResult
Reconfigure after losing readsNative re-execution (Tables R3 and R15)A single unavailable group affected the first certificate in 2,183 scenarios. The complete family supplied a same-size alternative in 659 (30.2%) and ruled one out in the other 1,524.
Identify a small set of sufficient readsCached scoring (Tables R1 and R2)With at most 96 evaluations per decision, FutureCert retained the fewest steps on average in 15/15 settings, including full-set fallbacks. On DreamerV3 Boxing it retained 9.7 steps versus 14.9 for LIME, the best comparator. Of 435 possible pairs of missing groups, the returned support avoided 206.8 pairs on average versus 116.9 for LIME.
Check whether a read budget sufficesCached scoring (Table R4)At both the 50% and 33% retention limits, FutureCert found a passing support most often in 15/15 settings. At the 50% limit its success rates ranged from 0.66 to 0.99 across settings, compared with 0.07 to 0.83 for the best comparator in each. At the 33% limit the ranges were 0.59–0.99 and 0.03–0.77.
Choose an imagination horizon per decisionReplay (Table R5)Truncating at the last selected step saved at least as many steps as a fixed 25-step horizon in 4/13 settings. In those four, replay saved 17–28% of steps and reproduced 81–92% of full-horizon decisions. The fixed horizon saved 17% and reproduced 69–90%.
Guide critic trainingExploratory training (Table R6)Giving selected steps more weight in the TD loss yielded higher mean returns over training than equal or shuffled weights on all three tasks. On Cup Catch the respective means were 844, 775 and 789 (100,000 environment steps, three seeds per task).

The native guarantee applies to the declared input, grouping, reference values, tolerance and planning stream. Changing the replacement rule requires a new check. The experiments in Tables R1, R2, and R4 rescore fixed trajectories. The replay in Table R5 uses the full trajectory to select a truncation point and measures neither runtime nor closed-loop return. Human interpretability remains a question for a user study.

When is the audit worth its cost? On 692 native MPPI inputs with nonempty minimum supports (Table R15), the first certificate, first four (or all if fewer), and complete family cost 445.2, 537.0 and 545.2 cumulative queries on average, including proof of minimum size. The first four covered 99.8% of single-group and 99.4% of double-group losses admitting a same-size solution. In a separate five-task study (Table R3), we audited all 4,095 non-full masks before storing the minimum supports. Among two-group loss scenarios admitting a valid support of any size, lookup among the stored minimum supports succeeded in 64–98% of those scenarios across tasks, compared with 6–98% for on-demand search with up to 32 queries. Both achieved 98% on Finger Spin. Prior auditing gained 28–60 percentage points on the other four tasks.

When can the same reads be reused? For reuse on fresh streams, we calibrated each frozen candidate pool on 64 independent streams per input, targeting an 80% pass probability at 95% confidence. The expanded candidate pool yielded a non-full certificate on 48/50 Finger Spin inputs, with 96.6% success on independent test trials. Coverage was 1/50 on Cheetah Run and 0/50 on Cup Catch. Non-full certification was rare on both tasks under both candidate pools.

Which tasks merit a closed-loop compression test? Development data predicted 70.3% read savings on Finger Spin. Closed-loop execution saved 50.9%. Both figures were at most 4.1% on each of the other four tasks. At comparable runtimes on Finger Spin, skipping reads and reducing planning iterations from six to four cost 0.77 and 0.79 times full planning time, but reduced return by 34.6 and 1.9, respectively.

The revision brings the five applications together in Table 4. Sections 5.4 and 5.5 report substitution and calibration results, and Table 5 compares predicted and measured read savings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.