acceptodds
Under review as a conference paper at ICLR 2027

Anisotropic Representations Improve Planning in JEPA World Models

Abstract

Latent world models learn action-conditioned dynamics in representation space and often score candidate actions by Euclidean distance to a goal representation. Joint training typically regularizes the representation to prevent collapse, but the resulting representation geometry also determines how terminal errors are weighted during planning. We show that accurate prediction and noncollapsed representations do not guarantee a task-aligned latent planning cost: isotropic Gaussian regularization can induce a geometry that ranks feasible outcomes differently from the task cost. To address this mismatch, we introduce AnisoWM with Reg, which replaces the fixed isotropic Gaussian target with a learnable diagonal covariance under fixed-trace and anisotropy constraints. The prediction objective, predictor architecture, and Euclidean planner remain unchanged; the target is used only during training. Our analysis characterizes the prediction-driven allocation of target variance, its dependence on the training distribution, and the conditions under which the induced metric reduces planning regret. Across four visual control environments, AnisoWM improves planning success over LeWorldModel in all four. Its latent planning cost also shows better agreement with task outcomes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
JEPA, World Model, Robotics

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_iR7eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper argues that traditional isotropic Gaussian regularization (SIGReg) in JEPA-based algorithms is not ideal because, when the learned representation is used for planning, the planner, when forced to choose between different directions, might prioritize low-variance directions over directions that arguably could take the agent closer to the goal state. The paper then introduces AnisioWM that “replaces the fixed isotropic Gaussian target with a learnable diagonal covariance under fixed- trace and anisotropy constraints”.

Critical strengths
  • Overall, the paper’s key point is very clear, and I find the argument convincing.
  • The theoretical results are detailed and appear correct.
  • Some results, such as the visualization in Figure 3, are quite didactic.
Critical weaknesses
  • Although the paper's main point seems clear, some of the jargon makes it difficult to follow some parts of the paper. An illustrative example is the sentence: “In a linear Gaussian setting, as process noise vanishes, the joint objective selects an approximately whitened representation, so Euclidean latent distance induces inverse-state- covariance weighting in state space. This places greater emphasis on low-variance directions and can change the ordering of feasible outcomes relative to the task cost, leaving positive planning regret even with exact conditional-mean prediction in representation space.”
  • I struggle to understand the point of Section 2. The manuscript barely interacts with its content, and it mostly seems like a dump of citations that are barely discussed in light of the paper’s claimed contributions.
  • *For the theoretical results, at the end of Proposition 1, the reduced objective is η2tr(KR)+λBRN(K)\frac{\eta}{2} \text{tr}(KR) + \lambda_B\mathcal{R}_N(K). You then set η→0\eta \rightarrow 0. That naturally leaves only the second term, which is SIGReg. But isn’t this result somewhat tautological? If you set one of the terms to zero, obviously the other one will dominate the solution, no? In this sense, the prediction itself is meaningless (and perfect). How realistic is this?
  • *The empirical results seem quite weak. For example, in Figure 2, each number is the average over three seeds, and there’s no sense of performance spread. How meaningful is it that after all this trouble one can improve over LeWM by no more than 6% (sometimes 1%)? What is the variance in performance for LeWM and AnisioWM?
Questions for authors
  • Please respond to the two weaknesses above highlighted with an asterisk.
  • It is said that a regularization weight of 0.09 was used. Was this hyperparameter tuned for AnisioWM? What were the values considered? I couldn’t find this information in the Appendix, but I might have missed it as well.
  • For LeWM, it is said that a regularization weight of 0.09 was used too. Why was this value selected? Was it tuned for LeWM? Was it tuned for * AnisioWM, and for consistency, was it also used in LeWM?
  • For the theory, the paper considers an affine encoder. But is this what was implemented? How do I reconcile this type of encoder with the common practice of using neural networks, for example?
  • On Eq. (10), what is v_i? Where is it used afterward? Was it defined?
8 comments
AuthorsOpenReview ↗

Response to Reviewer iR7e (1/4)

We thank the reviewer for the comments on the presentation, theoretical analysis, and empirical evaluation. We have revised the explanation of the planning mismatch and related work, and conducted additional experiments on finite-noise training, nonlinear representations, and performance variability.


W1: The explanation of the planning mismatch is difficult to follow.

A. Following the reviewer's feedback, we simplified the discussion in the Introduction to explain the planning mismatch without introducing the technical terminology prematurely.

The revised passage reads:

Our linear Gaussian analysis shows that the learned representation can give greater weight to state directions that vary less in the training data. When the goal cannot be reached exactly, this weighting can make the planner select a higher-task-cost outcome even with exact conditional-mean prediction.

Section 3 first introduces the state covariance and the metric induced by isotropic feature covariance (Section 3.1), then analyzes the geometry selected by joint prediction–SIGReg training under a process-noise model (Section 3.2).


W2: The Related Work section does not clearly connect prior methods to our contribution.

A. We revised Section 2 to explain how prior work relates to the planning mismatch studied in this paper. In particular, LeWM combines Euclidean latent-distance planning with SIGReg's isotropic Gaussian target. Our analysis shows that this combination can induce a planning cost that disagrees with the task cost, even with accurate prediction. Unlike approaches that modify prediction or introduce task-specific planning costs, AnisoWM addresses this mismatch by relaxing the isotropic regularization target while retaining LeWM's prediction objective and Euclidean planner.

AuthorsOpenReview ↗

Response to Reviewer iR7e (2/4)


W3: The small-noise result appears tautological because prediction loss vanishes.

Role of prediction in the analysis. We understand the reviewer's concern that the result may appear to follow simply from removing the prediction term as process noise approaches zero. However, our analysis does not assume zero prediction loss. We optimize the full prediction-regularization objective at positive process noise and characterize its global minimizers as the noise decreases.

For a linear encoder AA, the prediction loss decomposes as

Lpred(A,p)=12E[∥p(Ax,a)−Amη(x,a)∥2]+η2tr⁡(AWA⊤),\displaystyle \mathcal L_{\mathrm{pred}}(A,p) = \frac12\mathbb E\left[\Vert p(Ax,a)-Am_\eta(x,a)\Vert^2\right] + \frac{\eta}{2}\operatorname{tr}(AWA^\top),

where mη(x,a)m_\eta(x,a) is the conditional next-state mean and ηW\eta W is the process-noise covariance.

The first term measures error relative to the conditional mean. The second is the irreducible prediction error, which remains positive at η>0\eta>0 for a noncollapsed encoder and depends on the representation.

Importantly, Proposition 1 does not assume that the encoder is invertible or that the predictor is exact. It establishes that, for sufficiently small positive noise, every global minimizer has a noncollapsed encoder and recovers the exact encoded conditional mean.

As noise decreases, isotropic regularization determines the leading-order representation geometry. Theorem 1 then establishes the consequence for planning: the induced latent metric can select a different feasible outcome from the task cost, leaving strictly positive planning regret even as prediction loss and encoded rollout error vanish.

Thus, accurate prediction alone does not guarantee that Euclidean distances in the learned representation correctly rank candidate plans.

Finite-noise joint training. To examine whether this mismatch persists away from the small-noise limit, we additionally trained the linear encoder and predictor jointly using the full objective at four positive noise levels. The table below reports mean planning regret over ten seeds.

Noise orderingη=0.1\eta=0.10.030.010.001Small-noise limit
Favorable0.4710.4980.5060.5120.500
Reverse0.5540.5260.5190.5150.500

The trained models exhibit planning regret close to the predicted limit across the tested noise levels. At the largest noise level, the process-noise variances are 10% and 40% of the corresponding state variances. The mismatch therefore persists beyond the nearly deterministic regime.

Prediction-driven spectrum learning. Prediction also plays a direct role when the target spectrum is learnable. Proposition 2 shows that the limiting allocation minimizes tr⁡(LR)\operatorname{tr}(LR), where RR is the state-whitened predictive-noise covariance.

The finite-noise experiments in Appendix B.3 further show that the learned allocation responds to the noise structure. Under favorable noise ordering, anisotropy reduces planning regret; reversing the noise ordering can instead increase regret. The learned spectrum is therefore influenced by prediction, while its effect on planning depends on the relationship between predictive noise and task geometry.

AuthorsOpenReview ↗

Response to Reviewer iR7e (3/4)


W4: The reported planning gains are small, and their variability is unclear.

A. As requested, we report the individual planning success rates and sample standard deviations for both LeWM and AnisoWM, together with paired uncertainty estimates.

Both methods are evaluated on the same 50 initial–goal pairs per task using identical evaluator and planner settings.

TaskLeWM: 3072 / 1234 / 42AnisoWM: 3072 / 1234 / 42LeWM mean ± SDAnisoWM mean ± SDMean Δ (pp) [95% CI]
TwoRoom84 / 86 / 8890 / 94 / 9486.0 ± 2.092.7 ± 2.3+6.7 [0.7, 14.7]
Reacher74 / 82 / 7890 / 88 / 8878.0 ± 4.088.7 ± 1.2+10.7 [0.0, 21.3]
PushT90 / 94 / 8898 / 96 / 9690.7 ± 3.196.7 ± 1.2+6.0 [0.0, 13.3]
Cube76 / 78 / 6878 / 80 / 8074.0 ± 5.379.3 ± 1.2+5.3 [−2.7, 14.7]

We estimate uncertainty using 20,000 bootstrap draws, jointly resampling the shared evaluation pairs and independently resampling training runs within each method.

AnisoWM achieves higher mean success on all four tasks. However, when both evaluation-pair and training-run variation are considered, TwoRoom has the clearest support for improvement. The intervals for Reacher and PushT reach zero at the reported precision, while Cube's interval includes zero.

We also distinguish these matched comparisons from the published LeWM reference. As the reviewer noted, the improvement on PushT is small against the published result, whereas the matched comparison with reproduced LeWM yields a larger difference. We have clarified the sources of these comparisons in Section 5.1.

The additional results and evaluation protocol are provided in Appendix F.1.

AnisoWM also introduces no additional planning-time components. The target variances are learned during training and discarded afterward, leaving the predictor architecture and Euclidean planner unchanged.


Q2 & Q3: The choice of λ=0.09\lambda=0.09 and its sensitivity are unclear.

Regularization-weight selection. We clarify that λ=0.09\lambda=0.09 was adopted from the original LeWM study (Maes et al., 2026, Appendix G, Figure 16), which reports its strongest PushT performance near this value.

We did not tune λ\lambda specifically for AnisoWM. Both AnisoWM and reproduced LeWM use the same regularization weight in the primary comparisons.

Sensitivity analysis. To further examine the choice of regularization strength, we evaluated both methods at seven weights on PushT. The comparison uses seed 3072 and the same 50 evaluation pairs, with the remaining settings unchanged.

λ\lambda0.010.050.090.0950.100.200.50
LeWM success (%)92929092928268
AnisoWM success (%)100969896949070

AnisoWM achieves higher success across all tested weights, indicating that the improvement is not confined to λ=0.09\lambda=0.09. Both methods show reduced performance at larger regularization weights.

The sweep uses one training run per setting and is intended to characterize sensitivity rather than training-seed variability. The weight-selection rationale and comparison are reported in Appendix F.1 (Table 6).

Reviewer_1x4FOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies a gap between accurate latent prediction and useful Euclidean planning costs. In a linear Gaussian setting, it shows that joint prediction-SIGReg training selects a metric approaching qΣ−1q\Sigma^{-1}, which can produce positive planning regret despite noncollapsed representations and vanishing prediction error. AnisoWM replaces the fixed isotropic target with a learned diagonal covariance constrained by its trace and condition number. The analysis characterizes the selected metric and conditions under which it improves or worsens planning. Experiments report higher mean success than LeWM on four visual-control tasks, with additional cost-ranking diagnostics.

Critical strengths
  1. The analysis connects the training objective to planning failure.

Proposition 1 and Appendix E.2 establish the selected metric over all global training optima, allowing singular encoders during optimization and then proving that the optima are invertible at sufficiently small noise. Theorem 1 and Appendix E.5 turn this into a feasible finite-horizon control example with positive limiting regret. This rules out collapse and inaccurate prediction as explanations for the failure. Handling the expected finite-batch statistic, including its covariance minimum at qIqI, also makes the argument more precise than assuming exact population whitening.

  1. The intervention is small, with useful evidence that it changes planning-relevant geometry.

ΛReg changes only the training regularizer; prediction and planning retain the original latent coordinates. The primary comparison uses a shared κ=2\kappa=2 and reports matched training settings. Table 1 provides evidence beyond success rates: representation-only ordering improves from 0.564 to 0.661 in TwoRoom and from 0.790 to 0.893 in Reacher. These results support a geometric effect before predictor rollout.

The distinction from related approaches is meaningful: SCALE uses state-distance supervision, while DA-LeWM adds action-supervision heads. AnisoWM instead studies what can be achieved through target-distribution learning alone.

Critical weaknesses
  1. The reliability of the reported planning gains is not established.

Figure 2 reports rounded mean improvements of approximately 6, 3, 1, and 5 percentage points. However, Table 4 provides per-seed results only for AnisoWM, not LeWM, and the main comparison lacks uncertainty estimates for the differences. Each seed uses the same 50 initial-goal pairs, so the evaluations should not be treated as independent samples across seeds. In PushT, the reported comparison is only 96.7% versus 96%. The means support a descriptive improvement, but the paper does not establish how reliably these differences persist across training and evaluation variation. Figure 7 also indicates that the ranking results correspond to a single seed. This limits the empirical case for consistent improvement across all four tasks.

  1. The experiments do not isolate learned allocation from simply permitting anisotropy.

The spectrum plots and κ\kappa sweep show that the target changes and that the bound matters, but they do not compare learned targets with a fixed anisotropic target under the same trace, condition number, and training budget. This distinction matters because Equation (13) selects variance using prediction noise, not task cost; Appendix C.4 explicitly shows a regime where every κ>1\kappa>1 worsens regret. The paper acknowledges these limits, so this is not a contradiction of its theorems. Nevertheless, the empirical evidence does not yet establish that adaptive variance allocation, rather than a more general change in regularization, explains the gains.

Questions for authors

a) Please provide LeWM’s three per-seed success rates and paired uncertainty estimates for the AnisoWM-LeWM differences, accounting for training seeds and reuse of the same initial-goal pairs. Which gains remain supported?

b) At $\kappa=24, how does learned Λ\Lambda compare with a fixed target having 96 variances of 4/3 and 96 of 2/3, with otherwise identical training?

c) Which seed and source policies/checkpoints produced Table 1’s candidate pools? Do its representation-only and predicted-cost gains persist across the three existing training seeds, with uncertainty measured over initial-goal pairs rather than treating all candidate comparisons as independent?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1x4F (1/3)

We thank the reviewer for the feedback on the reliability of planning improvements and the contribution of learned variance allocation. Following the suggestions, we reproduced LeWM over three training runs, evaluated paired uncertainty, compared fixed and learned anisotropic targets, and extended the ordering analysis across training seeds.


W1 & Q1: The reliability of the reported planning gains is not established.

Paired planning evaluation. Following the reviewer's suggestion, we report per-seed LeWM results and paired uncertainty estimates for the differences in planning success. Both methods are evaluated over three training runs on the same initial–goal pairs.

TaskLeWM: 3072 / 1234 / 42AnisoWM: 3072 / 1234 / 42Mean Δ (pp)95% CI (pairs)95% CI (pairs + runs)
TwoRoom84 / 86 / 8890 / 94 / 94+6.7[2.0, 12.7][0.7, 14.7]
Reacher74 / 82 / 7890 / 88 / 88+10.7[3.3, 18.0][0.0, 21.3]
PushT90 / 94 / 8898 / 96 / 96+6.0[2.0, 10.7][0.0, 13.3]
Cube76 / 78 / 6878 / 80 / 80+5.3[0.0, 11.3][−2.7, 14.7]

We compute 95% confidence intervals using 20,000 bootstrap draws. The pair-only analysis jointly resamples the shared evaluation pairs across models, while the pairs-and-runs analysis additionally resamples training runs within each method.

AnisoWM achieves higher mean planning success on all four tasks. The pair-only intervals support improvements on TwoRoom, Reacher, and PushT. When training-run variation is also considered, only TwoRoom retains a clearly positive interval, indicating greater uncertainty in the improvements on the other tasks.

We also distinguish the matched results from the published LeWM reference. As the reviewer noted, the improvement on PushT is small when compared with the published result, whereas the matched evaluation against reproduced LeWM shows a larger difference. The two comparisons use different baseline sources and are reported separately.

The full per-seed results and paired uncertainty estimates are provided in Appendix F.1.

AuthorsOpenReview ↗

Response to Reviewer 1x4F (2/3)


W2 & Q2: The experiments do not isolate spectrum learning from allowing anisotropy.

Fixed-target comparison. As suggested, we compared AnisoWM with a fixed anisotropic target and isotropic LeWM under matched training conditions.

The fixed target assigns variance 4/34/3 to 96 coordinates and 2/32/3 to the remaining 96, giving trace 192 and condition number 2. AnisoWM uses the same trace and anisotropy bound but learns the target during training. All methods retain trainable encoders and predictors and use the same architecture, optimization settings, regularization weight, and training budget.

The table below reports planning success (%) over three runs. Each entry gives seeds 3072 / 1234 / 42, followed by their mean.

TargetTwoRoomReacherPushTCube
LeWM (isotropic)84 / 86 / 88 (86.0)74 / 82 / 78 (78.0)90 / 94 / 88 (90.7)76 / 78 / 68 (74.0)
Fixed anisotropic94 / 92 / 94 (93.3)88 / 82 / 92 (87.3)92 / 96 / 94 (94.0)78 / 78 / 80 (78.7)
AnisoWM (learned)90 / 94 / 94 (92.7)90 / 88 / 88 (88.7)98 / 96 / 96 (96.7)78 / 80 / 80 (79.3)

The fixed anisotropic target improves mean planning success over isotropic LeWM on all four tasks. AnisoWM performs slightly better on Reacher, PushT, and Cube, while the fixed target performs better on TwoRoom. These results indicate that much of the improvement comes from allowing anisotropy itself, rather than learning the spectrum.

Distribution-dependent allocation. We additionally conducted synthetic experiments in Appendix D.2 to examine how the preferred variance allocation changes across data distributions. Under identical trace and condition-number constraints, different distributions favor different spectral multiplicities, and a mismatched fixed spectrum incurs higher regret. The learned target identifies the predicted multiplicity in most runs.

Scope of our contribution. Our central contribution is to identify a planning mismatch that can arise when isotropic Gaussian regularization is combined with Euclidean latent-distance planning. The resulting latent distances can rank feasible outcomes differently from the task cost, even with exact conditional-mean prediction. Allowing anisotropic target variances makes it possible to adjust this geometry.

Since the appropriate variance allocation is generally unknown, AnisoWM learns the target spectrum jointly with the encoder and predictor. The fixed-target results show that much of the improvement comes from allowing anisotropy itself, while spectrum learning provides a data-dependent way to select the allocation rather than requiring it to be specified in advance.

To reflect this distinction, we revised the motivation and second contribution statement in Section 1 (Introduction) and added an interpretation of the fixed-target results in Section 5.2 (Planning Performance).

Prediction-driven allocation and task alignment. The reviewer also raised an important distinction between the spectrum selected by predictive training and the geometry required by the task cost.

As shown in Eq. (13), the limiting allocation minimizes tr⁡(LR)\operatorname{tr}(LR), where RR is the state-whitened prediction-noise covariance. The allocation is therefore determined by predictive noise rather than the downstream task cost. As discussed in Section 4.3, task alignment additionally requires the selected spectrum to match the task-relevant geometry. This is not guaranteed for arbitrary noise structures or task costs.

To examine this dependence beyond the small-noise limit, we additionally trained the linear encoder, predictor, and target jointly under favorable and reverse process-noise orderings. The following table reports trained / numerical-reference planning regret at η=0.03\eta=0.03, averaged over ten training seeds (Table 3 in Appendix B.3).

Noise orderingκ=1\kappa=123 (cc)9 (c2c^2)15
Favorable0.498 / 0.4880.083 / 0.0770.000 / 0.0000.483 / 0.4960.874 / 0.883
Reverse0.526 / 0.5121.032 / 1.0251.283 / 1.2821.718 / 1.7231.823 / 1.829

Under favorable noise ordering, moderate anisotropy reduces regret, but the improvement disappears at larger bounds. Under reverse ordering, anisotropy increases regret. Both patterns closely follow the numerical references.

These results confirm the limitation identified by the reviewer: learning the spectrum does not guarantee better task alignment. Its effect depends on how the predictive-noise structure relates to the task geometry.

AuthorsOpenReview ↗

Response to Reviewer 1x4F (3/3)


Q3: The candidate-pool sources and cross-seed ordering reliability are unclear.

Candidate-pool construction. The reviewer asked which seeds, policies, and checkpoints generated the candidate pools used in Table 1.

The original Table 1 and ranking figure (Figure 7 in the submission, Figure 8 in the revision) compare AnisoWM and reproduced LeWM at training seed 1234.

The candidate pools contain recorded planning trajectories from the anisotropy sweep at seed 3072, additional TwoRoom sweep configurations, and κ=2\kappa=2 runs with seeds 1234 and 42. The evaluation and planner seed is fixed at 42. Trajectories generated by either final checkpoint being compared are excluded, and both models evaluate the same remaining action sequences and realized outcomes.

We retain sequences reaching the task-specific diagnostic horizon and initial–goal pairs containing at least eight candidates with differing outcomes.

The table below summarizes the filtered candidate pools. The diagnostic horizon is distinct from the CEM planning horizon, and the candidate counts give the median and range across retained initial–goal pairs (out of 50 per task).

TaskHorizonRetained pairsCandidates per pair: median (range)
TwoRoom204596 (8–147)
Reacher255089.5 (28–102)
PushT254840.5 (8–95)
Cube1533100 (8–102)

Ordering across training seeds. Following the reviewer's suggestion, we extended the diagnostic to all three training runs and computed paired uncertainty over initial–goal pairs.

We evaluate two costs on the same recorded sequences. The representation-only cost JencJ_{\rm enc} uses the encoded terminal observation actually reached by each sequence, while JpredJ_{\rm pred} uses the predicted terminal embedding and includes predictor rollout. Both are compared against the ordering of the realized task outcomes.

The table below reports AnisoWM minus reproduced LeWM ordering accuracy on the 0–1 scale.

TaskCostΔ: 3072 / 1234 / 42Mean Δ [95% CI]
TwoRoomJencJ_{\rm enc}+0.040 / +0.097 / +0.028+0.055 [0.026, 0.085]
JpredJ_{\rm pred}+0.018 / +0.071 / +0.027+0.039 [0.025, 0.054]
ReacherJencJ_{\rm enc}+0.105 / +0.103 / +0.044+0.084 [0.061, 0.110]
JpredJ_{\rm pred}+0.044 / +0.048 / +0.003+0.031 [0.006, 0.056]
PushTJencJ_{\rm enc}−0.005 / −0.005 / 0.000−0.003 [−0.009, 0.002]
JpredJ_{\rm pred}+0.034 / +0.028 / −0.018+0.015 [−0.007, 0.037]
CubeJencJ_{\rm enc}+0.005 / +0.021 / +0.019+0.015 [−0.025, 0.061]
JpredJ_{\rm pred}+0.004 / +0.016 / +0.003+0.008 [−0.002, 0.020]

Ordering accuracy is computed within each initial–goal pair and averaged equally across pairs and training runs. The confidence intervals use 10,000 paired bootstrap draws over initial–goal pairs, keeping candidate comparisons within each pair together.

Both ordering metrics improve consistently on TwoRoom and Reacher, with confidence intervals supporting positive differences. Cube shows smaller, less conclusive improvements. On PushT, representation-only ordering remains nearly unchanged, while predicted-cost ordering improves in two runs but decreases in the third.

Thus, the ordering improvements are not equally consistent across tasks, even though mean planning success is higher on all four. Table 1 and Figure 8 identify the original seed-1234 comparison, while Appendix F.2 reports the three-run results and full evaluation protocol.

Dear Reviewer 1x4F,

Thank you again for your detailed feedback. Following your suggestions, we added matched three-run evaluations with paired uncertainty estimates, fixed-target comparisons, and a more detailed account of the candidate-pool construction and ordering results across seeds.

As the discussion period draws to a close, we would be grateful to know whether our responses have addressed your concerns regarding the reliability of the results and the scope of our contribution. We would be happy to clarify anything that remains unclear.

Reviewer_SP8FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how SIGReg, the anti-collapse regularizer in LeWorldModel (LeWM), shapes the Euclidean latent cost used for CEM planning. The analysis uses a linear-Gaussian model. In it, an isotropic feature covariance induces an inverse state-covariance metric, and joint prediction–SIGReg training selects this geometry as process noise vanishes. The resulting planner can incur positive finite-horizon regret despite exact conditional-mean prediction, and a nonlinear MLP toy illustrates this separation. AnisoWM/Λ\LambdaReg replaces the isotropic target with a learnable diagonal Gaussian covariance, under a fixed trace and a condition-number bound κ\kappa; the prediction loss, predictor and planner are unchanged. The analysis shows that the learned allocation is selected by the whitened process noise, and that in constructed systems it can cancel the inverse-covariance weighting. With κ=2\kappa=2, AnisoWM improves LeWM's planning success by 1–6 points in TwoRoom, Reacher, PushT and Cube (means over three seeds). The paper also reports latent-cost ordering diagnostics, local cost visualizations, the learned spectra and a single-run κ\kappa sweep.

Critical strengths
  1. The analysis gives a concrete account of how regularization shapes planning geometry. In the linear-Gaussian setting, joint prediction–SIGReg training selects an inverse state-covariance metric as noise vanishes (Proposition 1). Theorem 1 constructs systems with positive limiting regret at every fixed finite horizon despite exact conditional-mean prediction and vanishing rollout error. The analysis accounts for the finite-batch statistic and explicitly identifies cases where the learned target hurts, making its scope and limitations clear.

  2. Λ\LambdaReg is a small training modification with no additional planning-time computation: the learned target is discarded after training, and the predictor architecture and planner are unchanged. At a shared κ=2\kappa=2, both planning success and JpredJ_{\text{pred}} ordering accuracy improve in all four environments (Fig. 2; Table 1). The method is easy to adopt, so even moderate gains could be useful if their robustness is established.

Critical weaknesses
  1. The contribution of learning the variance allocation is not isolated. Fixed anisotropic targets and a SIGReg weight sweep are missing, so the comparisons do not distinguish adaptive allocation from benefits of allowing anisotropy or adjusting regularization strength. The theory characterizes global optima in a small-noise regime; whether finite-budget gradient-based joint training realizes the analyzed geometry and planning benefits at finite noise remains unexamined. Section 3.4 demonstrates SIGReg's failure but does not evaluate Λ\LambdaReg. Thus, the empirical gains are encouraging, but their attribution to the learned allocation and its proposed mechanism remains incomplete.

  2. The reporting is insufficient to assess the robustness and generality of the empirical improvements. The 50-pair, three-seed comparison lacks per-seed LeWM results and paired uncertainty estimates. The single-run κ\kappa sweep cannot establish reliable differences between settings, and it is not stated whether κ=2\kappa=2 was chosen before the sweep. Table 1 appears to use one seed, while Figures 3 and 8 lack aggregate local-geometry statistics. These diagnostics evaluate different comparisons and need not agree; their representativeness and stability across seeds nevertheless remain unclear.

Questions for authors
  1. In the linear setup of Appendices B.3 and C, could you evaluate Λ\LambdaReg at finite noise and with gradient-based joint training of the encoder, predictor and target under a finite training budget? Please compare the learned allocation and regret with the theoretical optima across noise levels and κ\kappa, including κ=c=cond(Σ)\kappa=c=\mathrm{cond}(\Sigma) and κ>c2\kappa>c^2 under the favorable noise ordering, and the reverse ordering. The purpose is to assess practical optimization and the applicability of the small-noise limits, without requiring exact limiting behavior at finite noise. Adding Λ\LambdaReg to the nonlinear Section 3.4 toy would further test whether it repairs the observed separation.

  2. With matched training budgets, how does AnisoWM compare with a fixed two-level target at κ=2\kappa=2, the final learned spectrum held fixed throughout a fresh training run, and LeWM with λ\lambda varied around 0.09? In the expected linear-model objective, fixing a global optimum's target spectrum still permits the encoder to attain that geometry by rotation. The comparison would test whether learning the allocation improves neural-network training beyond simply allowing anisotropy or adjusting regularization strength.

  3. Could you provide per-seed LeWM results and per-pair outcomes for both methods, with paired uncertainty estimates accounting for variation across training seeds and reuse of the same initial–goal pairs? The paper motivates a shared κ=2\kappa=2; was this value fixed before the Table 6 sweep, and which seeds underlie that table? These details would allow the reliability of the reported gains to be assessed.

  4. How were the examples in Figures 3 and 8 selected, and what are the local Spearman correlations aggregated over evaluated goals and seeds, with uncertainty estimates? Please also identify the seed coverage of Table 1 and report the ordering diagnostic across seeds. Clarifying the sampled comparisons and latent costs used by the two diagnostics would help establish how broadly the illustrated geometric improvements hold.

5 comments
AuthorsOpenReview ↗

Response to Reviewer SP8F (1/4)

We thank the reviewer for the careful comments on target allocation, finite-noise training, and the reliability of the planning results. We conducted fixed-target controls, a matched regularization-weight sweep, finite-noise joint-training experiments, and additional evaluations across seeds and goals.


W1 & Q2: The contribution of learning the variance allocation is not isolated.

Fixed and frozen targets. Following the reviewer's suggestion, we compared AnisoWM with isotropic LeWM, a fixed anisotropic target, and a frozen learned spectrum under matched training conditions.

The fixed target assigns variance 4/34/3 to 96 dimensions and 2/32/3 to the remaining 96, giving trace 192 and condition number 2. AnisoWM uses the same trace and anisotropy bound but learns the target during training. For the frozen-spectrum control, we hold the spectrum learned by AnisoWM fixed and retrain a fresh encoder and predictor. All methods use the same architecture, optimization settings, and training budget. We evaluate three training runs on the same 50 initial–goal pairs per task. Full per-seed results are provided in Appendix F.1 (Table 7).

TargetTwoRoomReacherPushTCube
LeWM (isotropic)86.078.090.774.0
Fixed anisotropic93.387.394.078.7
Frozen learned spectrum91.386.796.078.0
AnisoWM (learned)92.788.796.779.3

Both fixed and frozen anisotropic targets improve mean planning success over isotropic LeWM. AnisoWM provides additional improvements over the frozen learned spectrum, although the differences are modest. These results indicate that much of the benefit comes from allowing anisotropy itself.

Regularization strength. As suggested, we also compared LeWM and AnisoWM across seven regularization weights on PushT to examine whether the improvements depend on the choice of λ\lambda. Both methods use seed 3072 and the same evaluation pairs.

λ\lambda0.010.050.090.0950.100.200.50
LeWM92929092928268
AnisoWM100969896949070

AnisoWM achieves higher planning success at every tested weight, indicating that the improvement is not confined to a particular choice of λ\lambda. The shared value λ=0.09\lambda=0.09 was adopted from LeWM rather than tuned for AnisoWM. This sweep uses a single training seed, and the results are reported in Appendix F.1 (Table 6).

Distribution-dependent allocation. We additionally conducted synthetic experiments in Appendix D.2 (Table 4) to examine whether the preferred variance allocation changes across data distributions. Under identical trace and condition-number constraints, two distributions favor different spectral multiplicities, and a mismatched fixed spectrum incurs higher regret. The learned target identifies the predicted multiplicity in most runs.

Scope of our contribution. Our central contribution is to identify a planning mismatch that can arise when isotropic Gaussian regularization is combined with Euclidean latent-distance planning. The resulting latent distances can rank feasible outcomes differently from the task cost, even with exact conditional-mean prediction. Allowing anisotropic target variances makes it possible to adjust this geometry.

Since the appropriate variance allocation is generally unknown, AnisoWM learns the target spectrum jointly with the encoder and predictor. The fixed-target results show that much of the improvement comes from allowing anisotropy itself, while spectrum learning provides a data-dependent way to select the allocation rather than requiring it to be specified in advance.

To reflect this distinction, we revised the motivation and second contribution statement in Section 1 (Introduction) and added an interpretation of the fixed-target results in Section 5.2 (Planning Performance).

AuthorsOpenReview ↗

Response to Reviewer SP8F (2/4)


Q1: The small-noise predictions have not been tested under finite-budget joint training.

Finite-noise joint training. Following the reviewer's suggestion, we jointly trained the linear encoder, action-conditioned predictor, and target at finite noise levels to examine whether gradient-based training recovers the geometry and planning behavior predicted by our analysis.

We use the linear system in Appendix B.3 under both favorable and reverse process-noise orderings, with Σ=diag(1.5,0.5)\Sigma=\mathrm{diag}(1.5,0.5) and c=cond(Σ)=3c=\mathrm{cond}(\Sigma)=3. Each setting is trained for 20k Adam steps over ten seeds. We compare the learned allocation and planning regret with numerical references obtained by minimizing the expected training objective. The full experimental protocol is provided in Appendix B.3.

The table below reports trained / reference planning regret at η=0.03\eta=0.03 (Table 3 in Appendix B.3).

Noise orderingκ=1\kappa=123 (cc)9 (c2c^2)15
Favorable0.498 / 0.4880.083 / 0.0770.000 / 0.0000.483 / 0.4960.874 / 0.883
Reverse0.526 / 0.5121.032 / 1.0251.283 / 1.2821.718 / 1.7231.823 / 1.829

Under favorable noise ordering, regret decreases as κ\kappa approaches cc but increases again for larger bounds. Under reverse ordering, anisotropy increases regret. At moderate noise levels, finite-budget joint training closely follows the predicted dependence of planning regret and target allocation on κ\kappa and noise ordering. At smaller noise levels and large anisotropy bounds, however, the fixed training budget can be insufficient to reach the predicted allocation. These cases illustrate the optimization limits of finite-budget training, as discussed in Appendix B.3.

Nonlinear validation. The reviewer also suggested testing whether Λ\LambdaReg improves the planning mismatch observed in the nonlinear experiment of Section 3.4. We therefore added Λ\LambdaReg to the same jointly trained MLP encoder and predictor, keeping the architecture, planner, and training budget unchanged.

The table reports mean normalized planning regret over ten training seeds.

Targetη=3×10−3\eta=3\times10^{-3}10−310^{-3}3×10−43\times10^{-4}10−410^{-4}
SIGReg (κ=1\kappa=1)0.1630.1650.1530.154
Λ\LambdaReg (κ=2\kappa=2)0.0630.0960.1130.117

Λ\LambdaReg reduces mean planning regret across all tested noise levels, with paired confidence intervals supporting the improvements. Increasing the anisotropy bound toward cond(Σ)\mathrm{cond}(\Sigma) further reduces regret, while the learned state-space geometry moves closer to the Euclidean task geometry, as illustrated in the revised Figure 2.

AuthorsOpenReview ↗

Response to Reviewer SP8F (3/4)


W2 & Q3: The reliability of the planning gains and the choice of κ\kappa are unclear.

Planning success and paired uncertainty. Following the reviewer's suggestion, we reproduced LeWM over three training seeds and evaluated it alongside AnisoWM on the same 50 initial–goal pairs per task. We report per-seed success rates and paired confidence intervals to examine the reliability of the planning improvements.

TaskLeWM: 3072 / 1234 / 42AnisoWM: 3072 / 1234 / 42Mean Δ\Delta (pp)95% CI (pairs)95% CI (pairs + runs)
TwoRoom84 / 86 / 8890 / 94 / 94+6.7[2.0, 12.7][0.7, 14.7]
Reacher74 / 82 / 7890 / 88 / 88+10.7[3.3, 18.0][0.0, 21.3]
PushT90 / 94 / 8898 / 96 / 96+6.0[2.0, 10.7][0.0, 13.3]
Cube76 / 78 / 6878 / 80 / 80+5.3[0.0, 11.3][−2.7, 14.7]

We estimate uncertainty using 20,000 bootstrap draws. The pair-only analysis jointly resamples the shared evaluation pairs, while the pairs-and-runs analysis additionally resamples training runs within each method.

AnisoWM achieves higher mean planning success on all four tasks. The pair-only intervals support improvements on TwoRoom, Reacher, and PushT. When training-run variation is also considered, only TwoRoom retains a clearly positive interval, indicating greater uncertainty in the improvements on the other tasks.

Per-pair outcomes. As requested, we also compared the number of successful runs for each shared initial–goal pair. The table counts pairs where AnisoWM achieves more successes, both methods achieve the same number, or LeWM achieves more successes.

TaskAnisoWM higherEqualLeWM higher
TwoRoom6440
Reacher20246
PushT7430
Cube8393

Most evaluation pairs have the same number of successes under both methods. Among pairs with different outcomes, improvements are more frequent than decreases on all four tasks. These counts complement the confidence intervals by showing how the differences are distributed across evaluation cases.

Choice of κ\kappa. The reviewer also asked whether the shared anisotropy bound was selected before the sensitivity sweep. We fixed κ=2\kappa=2 before conducting the sweep and did not tune it separately for each task.

The anisotropic settings in the sweep use training seed 3072, while the κ=1\kappa=1 reference is taken from the published LeWM results. Our independently reproduced LeWM results are reported above and in Appendix F.1. We therefore interpret the sweep as a single-run sensitivity analysis, rather than evidence that the selected bound is optimal across training seeds.

Reviewer_YH5SOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how isotropic representation regularization can produce a latent distance that ranks feasible outcomes differently from the task cost. AnisoWM replaces the fixed Gaussian target with a learned diagonal covariance constrained by its trace and condition number, while leaving the predictor and planner unchanged. A linear-Gaussian analysis, a nonlinear toy experiment, and four visual-control benchmarks support the approach.

Critical strengths
  • The paper isolates an important failure mode: accurate latent prediction does not ensure that Euclidean latent distance is a useful planning cost. Separating representation-only ranking from predicted-endpoint ranking in Table 1 helps identify whether differences arise before or during rollout.
  • The intervention is small and the primary comparison is well controlled. A shared anisotropy bound of 2 improves mean success over a retrained LeWM baseline in all four environments without changing the predictor or adding inference-time machinery.
Critical weaknesses
  • The additional value of learning the target spectrum is not isolated experimentally. Proposition 2 gives a prediction-driven selection rule, and the paper appropriately states the additional conditions needed for task alignment. Different learned spectra across environments do not by themselves show that learning the spectrum is necessary for the control gains. A fixed nonuniform spectrum with matched trace and condition number would test whether learning the eigenvalues adds value beyond permitting anisotropy; the encoder should remain trainable in both cases.
  • Some gains are small relative to the evaluation size. PushT moves from approximately 96% to 97%, with 50 initial–goal pairs per environment. Appendix F.1 gives individual AnisoWM seeds but not paired LeWM results, and the ranking diagnostic is shown for one seed. Uncertainty estimates for paired differences are needed to assess consistency. In particular, PushT’s representation-only ordering is nearly unchanged and slightly lower, even though its predicted-cost ordering improves.
Questions for authors
  1. At matched trace and condition number, how does the learned target compare with a fixed nonuniform spectrum while retaining the same trainable encoder and predictor? This would isolate the added value of learning the spectrum rather than merely allowing anisotropy.
  2. Please provide paired per-seed baseline results and uncertainty estimates for success-rate and ranking differences. Are the small representation-only differences in Cube and PushT consistent across seeds, and does PushT consistently improve only when predictor rollout is included?
3 comments
AuthorsOpenReview ↗

Response to Reviewer YH5S (1/2)

We thank the reviewer for the feedback on target allocation and the reliability of planning improvements. Following the suggestions, we extended the fixed-target comparison to three training runs per task and evaluated planning success and ordering accuracy against independently reproduced LeWM models.


W1 & Q1: The additional value of learning the target spectrum is not isolated.

Fixed-target comparison. As suggested, we compared AnisoWM with a fixed anisotropic target and isotropic LeWM across all four tasks, to separate the effect of allowing anisotropy from that of learning the variance allocation.

The fixed target assigns variance 4/34/3 to 96 coordinates and 2/32/3 to the remaining 96, giving trace 192 and condition number 2. The learned target has the same trace, starts isotropic, and reaches a final condition number of 2.00. Both anisotropic settings retain trainable encoders and predictors. Architecture, optimization, regularization weight, and training budget are unchanged. Each method is trained with three runs and evaluated on the same 50 initial–goal pairs.

Planning success rate (%) is reported below:

TaskLeWMFixed anisotropicAnisoWM (Ours)
TwoRoom86.094 / 92 / 94 (93.3)92.7
Reacher78.088 / 82 / 92 (87.3)88.7
PushT90.792 / 96 / 94 (94.0)96.7
Cube74.078 / 78 / 80 (78.7)79.3

All results are percentages. LeWM and AnisoWM entries are three-run means; fixed-target entries show runs with random seeds of 3072, 1234, and 42, followed by their mean in parentheses.

The fixed anisotropic target improves mean planning success over isotropic LeWM on all four tasks. AnisoWM achieves slightly higher success on Reacher, PushT, and Cube, while the fixed target performs better on TwoRoom. These results suggest that much of the improvement comes from allowing anisotropy itself, rather than learning the spectrum.

Distribution-dependent spectrum selection. We also added synthetic experiments in Appendix D.2 to examine how the preferred variance allocation changes across data distributions. Under identical trace and condition-number constraints, two synthetic distributions favor different spectral multiplicities, and a mismatched fixed spectrum incurs higher regret. The learned target selects the predicted multiplicity in 8–10 of 10 runs.

Scope of our contribution. Our central contribution is to identify a planning mismatch that can arise when isotropic Gaussian regularization is combined with Euclidean latent-distance planning. The resulting latent distances can rank feasible outcomes differently from the task cost, even with exact conditional-mean prediction. Allowing anisotropic target variances makes it possible to adjust this geometry.

Since the appropriate variance allocation is generally unknown, AnisoWM learns the target spectrum jointly with the encoder and predictor. The fixed-target results show that much of the improvement comes from allowing anisotropy itself, while spectrum learning provides a data-dependent way to select the allocation rather than requiring it to be specified in advance.

To reflect this distinction, we revised the motivation and second contribution statement in Section 1 (Introduction) and added an interpretation of the fixed-target results in Section 5.2 (Planning Performance).

AuthorsOpenReview ↗

Response to Reviewer YH5S (2/2)


W2 & Q2: The planning gains lack paired uncertainty estimates, and the ordering analysis uses one seed.

The original submission reported AnisoWM's planning success over three training runs, with random seeds of 3072, 1234, and 42, but did not provide the corresponding per-seed LeWM results or uncertainty estimates. The ordering diagnostic in Table 1 was reported with seed 1234. Following the reviewer's suggestion, we report per-seed LeWM results and paired uncertainty estimates for planning success in Appendix F.1, and extend the ordering evaluation to all three training runs in Appendix F.2.

Planning success. The table below reports per-seed success rates and their mean differences (AnisoWM minus LeWM), where positive values indicate improvements over LeWM, together with 95% confidence intervals.

TaskLeWM: 3072 / 1234 / 42AnisoWM: 3072 / 1234 / 42Mean Δ (pp)95% CI (pairs)95% CI (pairs + runs)
TwoRoom84 / 86 / 8890 / 94 / 94+6.7[2.0, 12.7][0.7, 14.7]
Reacher74 / 82 / 7890 / 88 / 88+10.7[3.3, 18.0][0.0, 21.3]
PushT90 / 94 / 8898 / 96 / 96+6.0[2.0, 10.7][0.0, 13.3]
Cube76 / 78 / 6878 / 80 / 80+5.3[0.0, 11.3][−2.7, 14.7]

AnisoWM achieves higher mean planning success on all four tasks. The pair-only confidence intervals support improvements on TwoRoom, Reacher, and PushT. When training-run variation is also considered, only TwoRoom retains a clearly positive interval, indicating greater uncertainty in the improvements on the other tasks.

We estimate uncertainty using 20,000 bootstrap draws. The pair-only analysis resamples the 50 initial-goal pairs jointly across the trained models. The pairs-and-runs analysis additionally resamples the three training runs independently within each method. The same evaluation cases are retained across methods in both procedures.

Ordering accuracy. We also extended the ordering diagnostic to all three training runs. For each evaluated initial-goal pair, LeWM and AnisoWM score the same recorded action sequences against their realized outcomes. We consider both the representation-only cost JencJ_{\rm enc} and the predicted cost JpredJ_{\rm pred} used for planning. The table below reports AnisoWM minus LeWM ordering accuracy on the 0-1 scale.

TaskCost3072123442Mean Δ95% CI
TwoRoomJencJ_{\rm enc}+0.040+0.097+0.028+0.055[0.026, 0.085]
JpredJ_{\rm pred}+0.018+0.071+0.027+0.039[0.025, 0.054]
ReacherJencJ_{\rm enc}+0.105+0.103+0.044+0.084[0.061, 0.110]
JpredJ_{\rm pred}+0.044+0.048+0.003+0.031[0.006, 0.056]
PushTJencJ_{\rm enc}−0.005−0.0050.000−0.003[−0.009, 0.002]
JpredJ_{\rm pred}+0.034+0.028−0.018+0.015[−0.007, 0.037]
CubeJencJ_{\rm enc}+0.005+0.021+0.019+0.015[−0.025, 0.061]
JpredJ_{\rm pred}+0.004+0.016+0.003+0.008[−0.002, 0.020]

We retain 45 / 50 / 48 / 33 initial-goal pairs for TwoRoom / Reacher / PushT / Cube, respectively. Ordering accuracy is computed within each pair and averaged equally across pairs and training runs. The 95% intervals use 10,000 paired bootstrap draws over initial-goal pairs, keeping all candidate comparisons within each sampled pair together. These intervals measure variation across evaluation cases for the trained models. On TwoRoom and Reacher, both ordering metrics improve consistently across seeds, with confidence intervals supporting positive differences. Cube shows smaller, less conclusive improvements. On PushT, representation-only ordering remains nearly unchanged, while predicted-cost ordering improves in two runs but decreases in the third. Thus, the ordering improvements are not consistent across all tasks.

Dear Reviewer YH5S,

Thank you again for your thoughtful feedback. In response to your suggestions, we added three-run fixed-versus-learned target comparisons, distribution-dependent spectrum experiments, and paired uncertainty estimates for planning success and action ordering.

As the discussion period draws to a close, we would be grateful to know whether these additions have addressed your main concerns. We would be happy to clarify any remaining questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers

We thank all reviewers for their careful evaluations and constructive feedback. Before addressing the individual comments, we would like to clarify the source of the LeWM baseline results and distinguish the published reference from our independently reproduced models used in the additional experiments.


Clarification on the LeWM reference

We clarify that we independently reproduced LeWM using its official implementation and used these reproduced models for our additional analyses and controlled comparisons. However, for the main planning success comparison, we reported the published LeWM results from Figure 6 of Maes et al. (2026): 87 / 86 / 96 / 74% for TwoRoom / Reacher / PushT / Cube, respectively. These values appear in Figure 2 of the original submission (Figure 3 in the revision) and serve as the κ=1\kappa=1 reference in Figure 5 of the original submission (Figure 6 in the revision).

Our original description did not clearly distinguish between the independently reproduced LeWM models used in our analyses and the published LeWM success rates used in the main comparison. We have clarified this distinction in Section 5.1 and the corresponding figure captions of the revised manuscript. We apologize for the ambiguity.

Despite our best efforts to reproduce LeWM using the official implementation and reported training configurations, our three training runs achieved mean success rates of 86.0 / 78.0 / 90.7 / 74.0%, below the published results on three of the four tasks. We found that this was not unique to us, observing many reports that the published scores are not perfectly reproduced on GitHub issues of the official repository [https://github.com/lucas-maes/le-wm/issues]. Nevertheless, recognizing this gap in reproduction, we decided to compare with the officially reported figures instead of our reproduced ones, which are weaker and potentially could overemphasize our contribution.

For the questions regarding reproduction of LeWM, however, we answer with our reproduced scores below. We used the official LeWM code and their training configurations including the random seeds (3072, 1234, and 42). All models were evaluated on the same 50 initial-goal pairs per task with identical evaluator and planner settings.

AuthorsOpenReview ↗

Rebuttal Summary

We sincerely thank all reviewers for their careful and constructive feedback. Revisions are highlighted in blue and red in the revised PDF. We summarize the main updates below.

1. Theoretical clarification and validation [iR7e, SP8F, 1x4F]

  • Clarified the prediction–planning mismatch, the role of prediction-driven variance allocation, and related work: Secs. 1–4
  • Added finite-noise, nonlinear, and reverse noise-ordering analyses and experiments: Secs. 3.4, 4.4; Apps. B.3, C.4

2. Additional ablation studies [YH5S, SP8F, 1x4F, iR7e]

  • Fixed, frozen, and learned anisotropic targets: App. F.1, Tab. 7
  • Distribution-dependent spectrum selection: App. D.2, Tab. 4
  • Regularization-weight and anisotropy-bound sensitivity: App. F.1, Tab. 6; Sec. 5.2, Fig. 6; App. F.5, Tab. 13

3. Statistical reliability [YH5S, SP8F, 1x4F, iR7e]

  • Matched three-seed planning evaluations, per-seed results, and paired bootstrap confidence intervals: App. F.1, Tabs. 7–9
  • Three-seed action-ordering results, replacing the original single-seed comparison: Sec. 5.2, Tab. 1; App. F.2, Tab. 11

4. Local planning geometry [SP8F, iR7e]

  • Distinguished illustrative examples from aggregate correlation results: Sec. 5.2, Fig. 4; App. F.3, Tab. 12
  • Added further qualitative and randomly selected examples, including cases with limited or negative improvements: App. F.3, Figs. 9–10

5. Baseline attribution and evaluation details [iR7e, SP8F, 1x4F]

  • Clarified published versus independently reproduced LeWM results: Sec. 5.1, Figs. 3, 6; App. F.1
  • Detailed candidate-pool construction and evaluation protocols: App. F.1–F.2

We are grateful for the reviewers' thoughtful suggestions and the time they devoted to evaluating our work. Their feedback has helped us improve the clarity of our contributions, strengthen the experimental evidence, and present the results more transparently. We hope the revised manuscript and additional analyses address their concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.