acceptodds
Under review as a conference paper at ICLR 2027

Epistemic Action Quotients: A Theory of Consequence-Based Action Abstraction

Abstract

When should physically different sensing actions be treated as the same? We argue that the answer depends on what the robot needs to learn rather than on how the actions move it. We introduce Epistemic Action Quotients (EAQs), which define action identity through task-conditioned epistemic consequences: two actions are exactly equivalent for a demand when they induce the same distribution over task-projected belief updates. For finite data, we use bounded-diameter operational partitions and establish a Lipschitz within-block regret bound together with a sufficient stability condition under bounded drift. In the tested HM3D and Habitat settings, epistemic identity repeatedly differs from motor similarity and changes with the demand; a sensor pan and a whole-body yaw fall into the same class in all 20 seeds. In a controlled planning study, evaluating one representative per class reduces the candidate set eightfold with no observed regret. We further instantiate Online EAQ, which predicts action consequences from locally available pre-action information and treats demand changes as mandatory refresh events. In an HM3D pilot, task-triggered refresh accounts for most of the improvement over static caching, while all 25 transitions satisfying the realized-branch analogue of preserve the reference partition.

Primary area:
learning theory
Keywords:
Action Abstraction, Epistemic Equivalence, Task-Conditioned Representation, Embodied Intelligence, Decision-Theoretic Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FYgnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses the problem of robot learning. To the best of my knowledge, it tries to solve this by trying to understand what are the best ways to group different potential actions. This is tested in HM3D and Habitat.

Critical strengths

None

Critical weaknesses
  • Hard to understand the abstract.
    • action identity through task-conditioned epistemic consequences
      • What does action identity mean?
      • What does task-conditioned epistemic consequences mean?
      • “for a demand when they induce the same distribution over task-
      • projected belief updates”, what does task-projected belief updates mean?
      • “epistemic identity repeatedly differs from motor similarity and changes with the demand; a sensor pan and a whole-body yaw fall into the same class in all 20 seeds“, what does fall into the same class mean?
      • “satisfying the realized-branch analogue of”, what does realized-branch mean?
  • Unclear beyond the example it gives
    • “can pan its camera, yaw its whole body, or take a small step forward”, what are other applications of this approach
  • I read the abstract and introduction multiple times and it has been very hard for me to understand the motivation of the work and problem setting (e.g. beyond the initial example, what are other example applications of the problem), the approach (e.g. it is unclear to me why a complex approach is needed when motor similarity seems to be able to handled easily with a cost-based optimizer and some heuristic costs with different motions).
  • I feel like for this task, there are much simpler methods. For example, by associating some budget of different actions (can be heuristic, where different motor travel distance is associated to some cost, yaw is some cost…), and then using a LLM planner can solve this. Or a cost-based optimizer and some heuristic costs with different motions.
  • It is hard to understand the approach, and how it solves the problem. I had a hard time understanding the first paragraph and what demand is and what task projection is. In the approach, to aid in understanding, it is always helpful to start with some intuitive description of the approach. For example, have some running example, and use that to ground key parts of the approach (for example, demand q, task projection).
  • Suspicion of AI-generated text that is not reviewed
    • “We separate task-triggered refresh from the three-predictor consensus while keeping the seed-8603 initialization fixed.”
Questions for authors
  • Please see above and clarify the abstract and motivation. It has been very hard for me to understand the motivation of the work and problem setting (e.g. beyond the initial example, what are other example applications of the problem), the approach (e.g. it is unclear to me why a complex approach is needed when motor similarity seems to be able to handled easily with a cost-based optimizer and some heuristic costs with different motions).
5 comments
AuthorsOpenReview ↗

[1/2] Abstract and Method Clarification

Thank you for identifying points in the abstract, motivation, and method that need clarification, and for suggesting a running example to explain the demand and task projection. We address related comments together below.

To W1 and Q1: Abstract and research question

Given the available information and the current task, we ask which sensing actions can serve as informational substitutes, and how much decision value is lost when one representative replaces the other actions in its group.

We have rewritten the abstract on page 1 around this question and the corridor example from the Introduction:

Physically different sensing actions can provide the same information. To check whether the corridor ahead is clear, a robot may gain the same evidence by panning its camera or yawing its whole body, while changing the task from the center to the left sector can change how the full action set is grouped. Which actions are interchangeable thus depends on the task, not on motor similarity. We introduce Epistemic Action Quotients (EAQ), which group sensing actions by their consequences for the current task: given the same pre-action information, two actions are exactly equivalent if they induce the same distribution of task-relevant belief updates. Exact equivalence preserves value for any utility that depends only on these consequences; for approximate groups and Lipschitz utilities, we bound the regret of keeping one representative per group by the within-group consequence distance, and the grouping remains stable under suitable separation and drift conditions. In Habitat, sensor pan and body yaw share a group in all 20 seeds although a motor-based grouping separates them, and the full partitions differ between the center and left tasks. In synthetic experiments, one representative per group reduces the candidates from 64 to 8 with no observed regret, whereas motor-based grouping at the same budget incurs a mean normalized regret of 0.33. Online EAQ predicts consequences before acting and refreshes groups when the task changes; in a small HM3D pilot, this refresh accounts for most of its gain in partition agreement over static caching. In closed-loop HM3D runs, the reference consequences of visited states are compressible, but the current Online EAQ implementation does not improve cumulative proxy utility or reduce model calls relative to evaluating all seven actions.

The revised abstract no longer relies on the five terms you identified. For reference, their meanings in the paper are:

  • Action identity: informational interchangeability for the current task, given the same pre-action information.
  • Task-conditioned epistemic consequences: the possible changes in task-relevant beliefs induced by an action and its resulting observation.
  • Task-projected belief update: the part of a belief change retained by MqM_q; for qCq_C, this is the change in the center-sector depth belief.
  • Fall into the same class: belong to the same operational group under the specified distance, threshold, and grouping procedure.
  • Realized-branch: based on sampled observation outcomes used for the reference diagnostic, rather than direct verification of the full consequence distributions.

To W5: Demand, task projection, and a running example

Section 3 now opens with the following example before the formal definitions (page 3):

Returning to the corridor example in the Introduction, consider a robot seeking depth information about the area ahead or to its left. The demand qq specifies which information is needed, and the task projection MqM_q retains the relevant part of the belief. In our implementation, qCq_C selects the center-sector depth belief and qLq_L selects the left-sector depth belief. The evaluated tasks concern sector depth information; navigation safety is not evaluated. Although sensor pan and body yaw involve different motions, they may have similar information consequences for the current task. EAQ uses these consequences to determine which actions can be grouped. When the task shifts to the left, the retained belief component changes, and the action groups must be reconsidered.

A roadmap immediately after the example identifies the roles of Sections 3.1–3.3: exact equivalence, approximate grouping, and regret/stability conditions. The added examples in Sections 3.1 and 3.2 (page 4) distinguish equality of consequence distributions from operational co-membership at a specified threshold and numerical tolerance.

The example also continues in Section 3.3 (page 5): for a fixed task, a positive separation margin and sufficiently small drift are sufficient for stability; switching tasks changes the projection and requires reconsidering the groups. Online EAQ does not test this condition at runtime.

AuthorsOpenReview ↗

[2/2] Motivation and Supporting Evidence

To W2 and W3: Other applications and motivation

Beyond the corridor example, EAQ may apply to next-best-view selection, where candidate viewpoints can provide similar information about a region, and experimental design, where measurement schemes can induce similar belief updates about a target parameter. These connect to the active perception and experimental design literature discussed in Section 2 (page 3). They are potential applications; the empirical evidence reported in this paper comes from controlled synthetic experiments and Habitat/HM3D settings.

To W3, W4 and Q1: Heuristic costs, optimizers, and LLM planners

Execution costs, such as movement distance or turning cost, can guide action selection. They do not by themselves establish whether two actions provide similar task-relevant information or bound the regret from merging them. EAQ provides this grouping criterion; a planner then selects an action from the representatives.

In Table 1 (page 7), sensor pan and body yaw are separated by the motor partition but belong to the same information-consequence group under both tested tasks in all 20 seeds at δ=0.05\delta=0.05.

Table 2 (page 8) and Appendix H (pages 19–20) report the following results in a strongly separated, controlled synthetic setting over 20 seeds:

MethodCandidates evaluatedMean normalized regret
All actions640
Motor grouping80.331
EAQ80

Both grouping methods use the same candidate budget and select representatives with the same task-independent cost proxy before evaluating task values. EAQ has no observed regret in any of the 20 seeds. These controlled results illustrate that motor-based grouping can discard valuable candidates at the same evaluation budget.

Our guarantees concern the defined consequence-only utility under the stated assumptions; execution costs and other constraints belong in the subsequent selection problem. We did not evaluate LLM-planner integration, and fewer candidate evaluations do not imply proportional savings in total runtime.

Section 5.4 (pages 9–10) and Appendix M (pages 29–33) now report a 30-step closed-loop evaluation with one-step selection. The visited states contain compressible reference structure, but the current Online EAQ implementation does not improve cumulative proxy utility or reduce model calls relative to Full-7. The common scorer still predicts all seven actions at every step. These results do not establish navigation success or safety.

To W6 and Ethics Concern: AI use and the experimental description

Page 10 of the original manuscript (page 11 of the revision) discloses the use of ChatGPT, Codex, and Claude for feedback on the conceptual framework, experiment design, software implementation, analysis scripts, manuscript writing, and figure preparation, and states the authors' responsibility for the manuscript.

The quoted seed-8603 sentence describes a controlled comparison: all variants start from identical partitions generated by the seed-8603 predictor, so initialization differences do not confound the comparison of refresh and predictor aggregation (Table 4, page 9; Appendix L.2, page 26). We have rewritten the sentence in Section 5.3 (page 8) as: "All ablation variants start from the same seed-8603 partition, so their differences reflect refresh and consensus rather than initialization." Task-triggered refresh with a single predictor reduces pairwise partition error (PPE) from 0.3971 (Static) to 0.3452, matching full Online EAQ on this metric. This supports the paper's attribution of the main improvement in partition agreement to task-triggered refresh.

Thank you for your feedback. We are happy to clarify any remaining points.

Dear Reviewer FYgn,

We hope this revision addresses your concerns. Please let us know if there is anything further we can clarify; we would be happy to discuss any remaining issues.

Reviewer_nt7YOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Epistemic Action Quotients (which I will refer to as EAQ for the rest of the review). It is a notion of action identity for sensing actions in which two actions are equivalent for a given task demand when they induce the same distribution over task-projected belief updates. The main contributions are the exact quotient together with an approximate version based on bounded-diameter partitions computed by complete linkage and three guarantees i.e., value equivalence for exact classes, a within-block regret bound for Lipschitz utilities and a stability condition under which a partition is unchanged between consecutive decisions if per-action drift is below half the separation margin, and an instantiation called Online EAQ that predicts action consequences from pre-action features with a small MLP and refreshes the cached partition at demand changes.

Critical strengths

The conceptual contribution is well-motivated and nicely formalized. The paper identifies that most of the existing action-equivalence notions compare actions by their effect on the world rather than by what they reveal, and the connection drawn to Blackwell's comparison of experiments situates the idea exactly: EAQ is Blackwell equivalence restricted to the current demand.

The authors are consistently candid about the limits of their evidence. I note this as a strength because it is the reason I as a reader can be confident about where the gaps are.

Critical weaknesses

The paper's central motivation is that a planner can safely evaluate one representative per class. However, this claim is supported only by the synthetic candidate-reduction study of Appendix H, in which from my understanding 64 actions fall into eight Gaussian classes with minimum center separation 4 and within-class standard deviation 0.5, a regime in which complete linkage at delta = 2 cannot fail. No downstream planning result is reported in HM3D or Habitat. Because the authors themselves report that the online partitions in the embodied pilot have a pairwise partition error of 0.345, I am not able to conclude that representative-per-class evaluation would be safe in any embodied setting the paper studies.

The stability theorem is presented as the justification for reusing a cached partition under a fixed demand, but the measurements indicate that the regime it describes is rarely present. Within a fixed demand, the reference partition changes on 41% of steps in the Q1 episodes and on 57.4% of transitions in the pilot, and the theorem's sufficient condition certifies only 25 of 162 pilot transitions. Since the theorem itself follows in a few lines from the triangle inequality and Kantorovich-Rubinstein duality, its weight in the paper rests on empirical applicability, and the data show that applicability to be marginal. I am therefore not convinced the stability analysis should count as one of the paper's main contributions.

Questions for authors
  1. Has representative-per-class evaluation been tested with an actual planner on HM3D or Habitat? If so, what regret did it incur relative to evaluating all seven actions?
  2. What ARI and PPE do trivial partitions (all singletons; a single block) achieve on the pilot suite?
  3. How sensitive is the Habitat pan/yaw result (Table 1) to δ? The value δ = 0.05 appears without justification.
6 comments
AuthorsOpenReview ↗

[1/2] Embodied Representative Selection and Regret

Thank you for recognizing our motivation, formalization and evidence boundaries. We first address embodied selection, then stability, trivial baselines and threshold sensitivity.

To W1 and Q1: Embodied representative selection

The submission did not evaluate a downstream embodied selector. The 64-to-8 candidate reduction with zero observed regret in Table 2 and Appendix H was synthetic; Appendix L.2 explicitly states that the embodied pilot evaluated partitions, not downstream representative selection.

We have now run a closed-loop experiment in the tested static, deterministic configuration: Full-7, Online EAQ and Static each run 30 steps from six original starts across three HM3D scenes, totaling 540 selections and 3,780 reference branches for post hoc evaluation. Each policy acts on its own evolving observations. The new utility is the one-step reduction in the depth-discrimination risk proxy

1−max⁡(pnear,pfar),1-\max(p_{\mathrm{near}},p_{\mathrm{far}}),

where near/far probability masses use fixed depth bins. This does not measure navigation success or safety.

We define reference-branch regret as

Rt=max⁡autref(a)−utref(at).R_t=\max_a u_t^{\mathrm{ref}}(a)-u_t^{\mathrm{ref}}(a_t).

Full-7 uses the common predictive scorer, so its reference regret need not be zero. On the same 180 states visited by Online:

SelectorMean reference-branch regret
Online EAQ0.038531
Full-7 shadow0.039117
Uniform mean over all same-K subsets0.034247

Online minus Full-7 is −0.000586: 21 states improve, 42 worsen and 117 tie. The subset baseline averages all subsets of size K using the same scorer. These are same-state shadow/subset diagnostics, with no additional closed-loop trajectories. Online does not outperform the subset baseline on mean regret.

Executed closed-loop results are:

PolicyMean cumulative proxy utility/episodeMean terminal risk
Full-70.1193530.078775
Online EAQ−0.0168600.208774
Static−0.0181350.203977

Lower local regret on differently visited states therefore does not establish better trajectory utility. These are descriptive results from six paired starts.

To examine compressible structure, we analyzed the saved Online states and filled missing predictions with frozen-model inference. No new trajectories were generated. All three groupings use δ=0.05 and the original numerical tolerance. Pruning loss compares the best reference action with the best retained representative, excluding scoring-selection error within the retained set.

GroupingMean KMean pruning lossMaximum pruning loss
A: current reference consequences3.4222220.0011360.016843
B: hypothetical current refresh2.2888890.0245620.194852
C: actual cache2.1666670.0251550.195896

A has K<7 at all 180 states, retaining 51.11% fewer candidates on average. This connects compressible embodied reference structure with substitution loss under the specified proxy. A uses reference distances for grouping and representative selection and retains more candidates than B/C. Its small loss is partly constrained by the TV-based construction. It is neither deployable-method performance nor independent validation of population-distribution guarantees.

B still has reference within-group diameter above threshold in 133/180 states; violations also occur at 7/12 actual refreshes. For this utility, a diameter violation need not cause positive pruning loss. These mismatches cannot all be attributed to cache age. B retains the original cache-dependent consensus rule and also updates representatives, so its difference from C does not separately identify predictor error, consensus, representative updates or cache effects. We did not propagate hypothetical caches along new trajectories. The small mean B−C difference also masks opposing episode-level effects.

In this new closed-loop evaluation, the original cache-update rule is unchanged, but the common downstream scorer predicts and scores all seven actions at every step; Online then selects within its representatives. Online uses 222 model batches versus Full-7's 180. Thus, fewer candidates have not yielded demonstrated model-computation savings or deployment speedup.

These results support compressible reference structure in the tested states while exposing limitations of the current learned implementation. They do not establish lossless embodied selection. Theorem 3.5 bounds substitution loss for Lipschitz, consequence-only utilities using block diameters under true consequence laws; predicted proximity alone does not establish this premise. We have added these results to Section 5.4 and Table 5 (pp. 9-10), the protocol and detailed results to Appendix M (pp. 29-33; Tables 22-25), and the limitations to Section 6 (p. 10).

AuthorsOpenReview ↗

[2/2] Stability, Trivial Baselines, and Threshold Sensitivity

To W2: Stability coverage and its role

We agree that coverage is limited. We present stability as a supporting sufficient condition, without claiming broad empirical applicability. Theorem 3.7 requires a fixed task, positive separation margin and sufficiently small actionwise consequence drift, 2ηt<ρt2\eta_t<\rho_t. A fixed task alone does not justify reuse; failing this condition does not imply a partition change.

Rechecking the original pilot's 162 fixed-task transitions, of which 93 change partition, gives:

Realized-branch criterionCoveredExact-zero driftNonzero driftChanges within coverage
2η < ρ25/1622410
B < ρ28/1622440

The 24 shared zero-drift cases are exact zeros, not rounding or tolerance effects. Their empirical reference distance matrices, reconstructed from saved consequence vectors, are identical. Unchanged partitions follow by construction under the fixed deterministic clustering rules. There are 35 zero-drift transitions overall; the other 11 have negative margins and fail both criteria. Nonzero-drift coverage is thus only 1 and 4 transitions. This is a reference-level applicability check, not independent validation of population-law premises.

These are post hoc diagnostics: Online EAQ does not estimate these quantities at runtime. The revised Introduction (pp. 2-3), Section 4 (pp. 5-6) and Figure 1 (p. 2) distinguish sufficient conditions from certification of unchecked reuse. Section 5.3, Figure 3 (pp. 8-9) and Appendix L.6 (pp. 28-29; Table 21, p. 29) give the full counts. Figure 3 shows 81 positive-margin, nonzero-drift transitions on the left and all 162 on the right, including 57 negative-margin transitions with 36 changes.

The main formal content remains task-conditioned action identity, its ideal/operational approximations, and conditional value-equivalence and substitution-loss results.

To Q2: Trivial partitions on the same pilot

We recomputed both baselines on the original 180 pilot decisions, calculating metrics per decision before averaging, as in Appendix L.2. These differ from the new closed-loop states.

MethodARIPPEMean K
All-singletons0.0000000.3558207.000000
All-one-block0.0944440.6441801.000000
Static0.2847700.3970902.444444
Online EAQ0.3367650.3452382.444444

All-singletons has no false merges but retains all seven actions. Its PPE is better than Static's; Online improves on it by only about 1.06 percentage points, with mean K=2.444444. We therefore interpret partition error alongside compression, treating neither alone as proof of downstream utility.

The reference is never all-singletons and is all-one-block in 17 decisions. Under the established ARI convention, identical degenerate partitions receive ARI=1, explaining the all-one-block mean of 17/180. This comparison and interpretation are now in Appendix L.4, Table 20 (p. 28).

To Q3: Pan/yaw sensitivity and the choice of δ

We re-partitioned saved data from the original 20 seeds for both tasks and resolutions at 12 thresholds: 0, 0.005, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.075, 0.10, 0.20 and 1.0. C and L denote the center and left tasks. Pan/yaw co-grouping counts:

δ256: C256: L512: C512: L
018/2020/2017/2015/20
0.00520/2020/2020/2020/20
0.0520/2020/2020/2020/20

Pan/yaw co-grouping is 20/20 in every condition at every tested positive threshold, but full partitions change. Pairs list center then left; counts are out of 20. "Different" counts seeds with differing center/left partitions.

δPixelsMean K (C, L)One block (C, L)Different
0.005256(4.95, 5.40)(0, 0)14
0.005512(5.05, 5.35)(0, 0)13
0.05256(2.45, 3.25)(3, 0)20
0.05512(2.45, 3.25)(3, 0)20
0.10256(1.80, 2.85)(10, 1)18
0.10512(1.85, 2.85)(10, 1)18

At δ=1, all partitions have one block and task differences disappear. Stable pan/yaw co-grouping does not imply invariant partitions. The δ=0 results retain the original numerical tolerance and do not prove equality of population laws.

At δ=0.05, all 40 seed–task pairs have identical full partitions across resolutions. Section 5.1 (p. 7) now restricts resolution invariance to δ=0.05.

δ=0.05 was a prespecified operational posterior-TV tolerance. We have no record of validation-based tuning or evidence of optimality; it is not a calibrated planning-loss guarantee. The retrospective scan does not justify the original choice. Appendix J.1 and Table 15 (pp. 22-23) explain this and report the full scan, including K, one-block frequencies and task differences.

Dear Reviewer nt7Y,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

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

The paper introduces Epistemic Action Quotients (EAQ), a framework for grouping sensing actions according to their task-conditioned informational consequences rather than motor similarity. It distinguishes exact equivalence, ideal approximate partitions, and operational complete-linkage partitions, establishing conditional guarantees for value equivalence, within-block regret, and partition stability. The authors also propose Online EAQ, which maintains partitions using learned consequence predictors, predictor consensus, and demand-triggered refresh. Synthetic and Habitat/HM3D experiments demonstrate task-dependent action groupings, controlled partition recovery, candidate reduction, and modest improvements in online partition agreement.

Critical strengths
  1. Meaningful problem formulation. Defining sensing-action equivalence through task-relevant informational consequences rather than motor similarity provides an intuitive and potentially useful abstraction criterion for active perception.
  2. Carefully scoped theoretical framework. The paper distinguishes exact equivalence, ideal approximate partitions, and operational complete-linkage partitions, with explicit assumptions and corresponding guarantees. These include exact value equivalence, a Lipschitz within-block regret bound, and a sufficient drift-based partition-stability condition (Section 3), alongside conditional prediction-error extensions in the appendix.
  3. Controlled empirical evidence. In synthetic settings with well-separated consequence classes, complete-linkage clustering recovers the generating partition and reduces 64 candidate evaluations to eight without observed regret (Table 2). Habitat/HM3D experiments further demonstrate that action groupings depend on the task demand and can differ substantially from motor-based groupings.
  4. Transparent analysis of online behavior. Table 4 separates the effects of demand-triggered refresh and predictor aggregation, including cases where simpler variants achieve better partition agreement. The stability diagnostic provides evidence consistent with the theoretical sufficient condition while clearly acknowledging that it is not a deployable runtime certificate.
Critical weaknesses
  1. Theory-implementation gap in consequence distances. The theoretical guarantees concern Wasserstein distances between population-level consequence distributions, whereas Online EAQ uses half-L1 distances between deterministic predicted belief updates. Proposition D.2 provides conditional bounds linking these quantities through uniform prediction-matrix errors, but these errors are not estimated or validated in the online pilot. Furthermore, Theorem 3.7 does not certify cached partitions between refresh events. Consequently, the applicability of the theoretical guarantees to the learned online procedure remains uncertain.
  2. Theoretical distinctiveness and comparative evidence. The demand-conditioned notion of action equivalence is meaningful, but its distinctiveness relative to existing experimental-equivalence and value-preserving abstraction frameworks could be articulated more sharply. Several guarantees rely on established probability-law, Wasserstein, and separation arguments. Moreover, the empirical comparisons primarily involve motor clustering, static caching, and internal variants, leaving the advantage over alternative task-aware grouping approaches unclear.
  3. Incomplete operational sensitivity analysis. Although the controlled studies examine threshold and sample-size sensitivity, the learned online setting lacks systematic evaluation of δ, refresh frequency, and predictor ensemble size. These choices are important because reference partitions can change substantially even under a fixed demand (Section 5.1), and different refresh policies may involve different accuracy-computation trade-offs.
  4. Limited online validation and practical utility. The Online EAQ pilot comprises only six episodes across three HM3D scenes (180 decisions). Of 12 demand-triggered refreshes, only four change the operational partition, and Online EAQ coincides with Static in two episodes (Appendix L.4-L.5). Moreover, performance varies substantially across scenes: Scene S3 achieves ARI 0.078 and false-merge proportion 0.434, raising concerns about reliability when consequence prediction is inaccurate. Although synthetic experiments demonstrate candidate reduction without observed regret, the online pilot does not establish preserved decision quality or end-to-end computational savings relative to evaluating all candidate actions.
Questions for authors
  1. Operational guarantees and prediction reliability. Proposition D.2 provides conditional bounds based on prediction-matrix errors, but these errors are not quantified in the online pilot. Can the authors assess their magnitude and relation to false merges, particularly in Scene S3, where ARI falls to 0.078 and the false-merge proportion reaches 0.434? Can the existing data help distinguish shared prediction bias from other causes of poor partition agreement?
  2. Comparative evidence and decision utility. Could the authors clarify the practical distinction between EAQ and existing task-aware or information-gain-based grouping approaches? If a complete embodied planner is outside the paper’s scope, could an analysis of existing counterfactual branches or a lightweight decision-quality proxy assess whether the learned partitions preserve useful action choices?
  3. Refresh and threshold sensitivity. How sensitive are online ARI, false merges, and class count to the threshold δ = 0.05? Could the authors compare demand-triggered refresh with a simple periodic or per-step refresh on the existing trajectories, including the additional prediction and partitioning costs?
  4. Consensus trade-off and practical significance. Table 4 shows that Refresh-mean achieves higher ARI (0.3671) than Online EAQ (0.3368), whereas Online EAQ reduces the false-merge proportion from 0.2616 to 0.2421. While the rationale for prioritizing false merges is reasonable, can the authors demonstrate whether this reduction translates into meaningful improvements in decision quality? How consistent is this trade-off across episodes and scenes?
3 comments

Thank you for reviewing our submission. Your Paper Summary captures the main content of our work: grouping sensing actions by their task-relevant informational consequences, and studying the corresponding value, regret, and stability guarantees.

While preparing a point-by-point response, we noticed that some of the content in the Critical Strengths, Critical Weaknesses, and Questions sections does not appear to correspond to our manuscript. For example, Probe2Act, the fixed 12-probe sequence, and the common-path estimator in Section 4.1 do not appear in our paper, so we are unable to respond to these points as currently written. Our paper studies EAQ, that is, equivalence and approximate grouping of sensing actions by their informational consequences under the current task (Section 3, pp. 3–5 of the revised manuscript).

Could you please help check whether these sections correspond to this submission, and add or update your comments on our paper? We greatly value your feedback and would like to respond accurately to your specific questions about EAQ. Once the review is updated, we will reply promptly to each point.

Thank you for your help and for the further discussion.

Dear Reviewer pYiY, We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Dear Reviewer pYiY,

We hope this revision addresses your concerns. Please let us know if there is anything further we can clarify; we would be happy to discuss any remaining issues.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief Summary of Revisions to the Manuscript

Dear Area Chair, Here is the summary of our clarifications regarding the reviewers' critiques.

First, we clarify EAQ's core contribution: defining when sensing actions are informationally interchangeable given the available information and task, and establishing conditional bounds on substitution loss. The rewritten abstract and the center/left-sector running example explain the task projection and distinguish exact equivalence from operational grouping (p. 1; Section 3, pp. 3-5). The original definitions, theorem statements, and proofs are retained.

Second, we add evidence addressing the empirical questions: trivial-partition baselines on the original pilot (Appendix L.4, pp. 27-28), a 12-threshold sensitivity analysis (Appendix J.1, pp. 22-23), and closed-loop evaluation with one-step selection, accompanied by matched-state comparisons and regrouping diagnostics (Section 5.4, pp. 9-10; Appendix M, pp. 29-33). The diagnostics identify compressible reference structure in the visited states. The current Online EAQ implementation, however, does not improve cumulative proxy utility or reduce model calls relative to Full-7.

Third, we clarify the role of stability and the limits of cache reuse. The complete audit of the original pilot's 162 fixed-task transitions separates exact-zero from nonzero drift and shows limited coverage beyond zero-drift cases (Section 5.3 and Figure 3, pp. 8-9; Appendix L.6, pp. 28-29). These are post hoc reference checks, and the evaluated policy does not test the sufficient conditions online. We have aligned the Introduction and cache-policy description with this distinction and updated the Conclusion and Reproducibility Statement (pp. 10-11).

These revisions preserve EAQ's formal contribution while making the empirical evidence and current implementation limits more explicit. We hope the clearer presentation and additional analyses help the reviewers and Area Chair assess the contribution and its scope.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.