acceptodds
Under review as a conference paper at ICLR 2027

Subspace-Restricted Planning under Action Compression in Latent World Models

Abstract

Latent world models plan in a learned space, and many of them compress the action before it reaches the predictor. We show that this compression leaves the planner partly blind. On a released Cube-manipulation checkpoint, 15 of 25 action dimensions never reach the predictor, so the planning cost is exactly constant along them. These directions still move the robot. A perturbation along them changes the environment while the model's imagined rollout stays at numerical noise, so no check built on the model can detect it. The planner itself produces this motion. Along a blind direction no candidate looks better than another, so a sampling-based planner executes random motion that it never selected. Raising the sampling budget 27x does not remove this motion. We propose pruned sampling, which restricts the sampler to the directions the cost depends on. One SVD of the frozen weights gives those directions. The restriction needs no training and provably cannot worsen the best achievable plan. On Cube it raises success by 6 points at the standard budget. At that budget, projecting only the executed action gives a gain not distinguishable from this one. At the largest budget tested, pruned sampling keeps its gain and that projection does not. The same restriction extends beyond the exactly blind directions. The spectrum of the second moment of the cost gradient measures how many directions the cost varies along. On four of the five checkpoints studied, this spectrum collapses far inside the visible subspace. Sampling above its spectral gap raises success by 10.7 points on Cube and by 12.3 points on Cube-double. On Cube-double, pruned sampling alone gives no measurable gain. The gains hold for four samplers on Cube and three on Cube-double. A preregistered dose-response test and the same spectrum separate the two effects behind these gains, executed noise and sampling dilution. They tell apart the checkpoints where a restriction helps from the ones where it does not. Project page: https://subspace-restricted-planning.github.io/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world models, model-predictive control, sampling-based planning, action representation, active subspaces

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_E9mLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper analyzes the effects of action subspaces that are unobservable to latent world models and proposes using SVD for constraining action sampling to the observable subspace. The authors of the paper show that sampling-based planners cannot infer the cost of actions along unobservable directions as these directions have no effect on the model’s predicted cost. This potentially can harm the performance of the system as the planner may select arbitrary actions in this unobservable subspace as they uniformly have equivalent cost under the latent model. They validate their approach, showing that by restricting the sampling to the observable subspace, the performance of the planner improves.

Critical strengths

The proposed methodology is straightforward, theoretically grounded, and computationally inexpensive to implement. Performing an SVD of the frozen action encoder correctly identifies the observable action subspace without requiring retraining or fine-tuning of the underlying world model. The analysis on the singular values provides additional insight into which action directions potentially have relatively low sensitivity in the learned representation, and therefore, have little effect on the predicted planning cost. Lastly, there are many existing libraries that efficiently perform the SVD on matrices. For these reasons, the proposed approach is easy to integrate into existing latent world-model planning frameworks with little computational and architectural overhead.

Critical weaknesses
  • The proposed methodology has limited improvements over baselines. While the proposed methodology demonstrates statistically significant results on the Cube planning benchmark (Table. 1), the method does not provide the same statistically significant result across the other benchmarks. This limits the generalizability and practical utility of the proposed approach.
  • The proposed methodology has limited improvement over execution-only projection. The results in Section 5.2 demonstrate that projecting the selected action into the observable subspace before actuation performs comparably and sometimes better than the proposed subspace-aware sampling method. In cases where the proposed method performs better than the execution-only projection, the results are not statistically significant (many research fields use p <= 0.05). This questions the practical utility of the proposed sampling procedure over the simpler post-optimization projection procedure when it achieves comparable performance.
Questions for authors

Given that the improvement of pruned sampling over execution-only projection is not statistically significant, can the authors provide stronger evidence that doing this pruned sampling will be more beneficial than the simple execution-only projection.

1 comment

Thank you for the review, and we ran the experiment your questions asked for. The result supports your reading instead of our expectation, and we report it in full below.

Q / W2. Is there stronger evidence that pruned sampling is more useful than execution-only projection?

We do not have such evidence. After the new experiment, we no longer expect it at the standard budget.

The submission compared the two operators only on the visible subspace of W. Its larger gains come from the top eigendirections of the cost-gradient second moment C, where the paper tested only in-loop sampling. We have now run execution-only projection onto those directions, with the ambient CEM planner unchanged. We preregistered the prediction that in-loop sampling would be better on Cube-double. The prediction failed.

300 paired episodes, six seedsprojection − ambient (new)in-loop − ambient (paper)in-loop − projection
Cube (k = 4)+8.7 [+5.3, +12.3]+10.7 [+6.7, +15.0]+2.0 [−1.7, +5.7], p = 0.36
Cube-double (k = 3)+13.3 [+8.7, +18.3]+12.3 [+7.3, +17.3]−1.0 [−4.0, +2.0], p = 0.66

On both checkpoints, projection alone gives a gain that is not distinguishable from the in-loop gain. Two comparisons in the paper lean toward the in-loop sampler: +2.3 on Cube at N = 2700, and on PushT a loss of 8.0 points for projection against 2.0 in-loop. Neither difference is resolved (p = 0.12 for both). We will not claim that the sampler is better than the projection.

We believe this result raises the practical value of the paper, for two reasons.

  • The projection is one matrix product at actuation, the simplest form of the method. On Cube-double it raises success from 32.7% to 46.0%, with all six seeds positive.
  • The gain depends on which subspace is used, and finding that subspace is what the paper's analysis provides. Projection onto the visible subspace of W gains +1.3 on Cube-double (600 paired episodes, CI [−0.7, +3.5]). Projection onto the three directions that the spectrum selects gains +13.3. The paired difference between the two projections is +10.7 [+5.7, +15.7], with all six seeds positive. The execution projection is itself a construction of this paper (Figure 2), and Theorem 4.3 is the reason it loses nothing on the blind subspace.

The revision makes two changes that follow from your question.

  1. The execution projection becomes a method on equal footing with the sampler, with the table above in the main text.
  2. We withdraw the paper's explanation of the Cube-double gain as sampling dilution (abstract, Section 5.3). The projection does not change the search, so the gain cannot come from the search. The data support a different account. The gain comes from removing the component of the executed action along directions where the cost is flat. This is consistent with Section 5.2. On Cube, the sampling scale contracts almost equally along visible (×0.37) and blind (×0.41) directions, so selection along the visible directions is weak as well.

W1. Limited improvement across benchmarks.

We understand this reading, and our Table 1 caused it. It shows only Cube and PushT, and the second main result is only in Section 5.3. On Cube-double the task tier gains +12.3 points (p < 0.001, all six seeds positive), and the gain holds under MPPI (+16.7) and random shooting (+14.0). The full picture at the standard budget is:

checkpointambient (%)sampling, visible subspaceprojection, visible subspacesampling, top-k of C
Cube63.0+6.0 (p = 0.005)+6.7 (p = 0.0002)+10.7 (k = 4, p < 0.001)
Cube-double32.7+0.0 [−3.7, +3.7]+1.3 [−0.7, +3.5]+12.3 (k = 3, p < 0.001)
Scene77.3−3.7 [−8.0, +0.3]−1.0 (p = 0.75)n/a (no spectral gap)
PushT88.0+1.3 (p = 0.82)n/a (full rank)−2.0 (k = 2, p = 0.66)
Two-Room87.3+1.3 (p = 0.75)n/a (full rank)+0.0 (7 flips each way)

Two checkpoints gain more than 10 points. The other three show no significant change. We agree that the method does not help everywhere, and we do not want to present it as if it did. The paper reports where and why. On Scene the spectrum has no gap, and the criterion says not to cut. On the two full-rank checkpoints the restriction costs nothing measurable. The architectural tier provably cannot lower the best achievable plan (Proposition 4.2), and its measured change is never significantly negative. On Cube the task-tier gain holds for four samplers: +10.7 (CEM), +20.0 (MPPI), +10.7 (shooting), +18.7 (iCEM). On PushT and Two-Room (full rank) the first column prunes the weakest singular directions (Section 4.3). The revision merges these rows and the new results into the main table.

Your question improved the paper. If these answers address your concerns, we would be grateful if you would reconsider your assessment, and we are glad to answer any further question.

Reviewer_8MwHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method to mitigate action blindness during planning with a world model. The problem is that when the action embedding layer is a projection to a lower-dimensional space, the model becomes blind to some of the action dimensions, which makes downstream planning with CEM basically a random walk along the ignored dimensions. They propose two different methods for this, but both essentially ignore the action dimensions with small eigenvalues. Interestingly, simply ignoring the action dimensions that are likely to be random-walking works quite well.

Critical strengths

None

Critical weaknesses
  • Presentation is hard to follow: The manuscript is very hard to follow, and it is difficult to grasp the core algorithm. It took me a while to figure out what this paper is actually doing. The figures are not very helpful either; the text in them is too small to read, and none of the figures visualizes the algorithm or the problem that this paper targets.
  • Artificial motivation: The first thing that came to my mind was: why not simply widen the action embedding layer? I think this is an important question, since it would be a far simpler solution to the problem the paper targets. If the problem resides in the world model’s implementation, why should we mitigate it at the planning stage? However, I could not find any experiment related to this in the main paper or the supplementary material.
  • Also, related to the second point, I wonder whether the raw actions are full rank. The design of down-projecting actions is based on the manifold hypothesis: actions are not full rank and reside in some low-dimensional subspace, which is generally true. However, I could not find any experiment showing the effective rank of the raw action data.
  • Limited experiments: Experiments are only conducted in simple synthetic environments, without real-world experiments or more challenging settings with rich contact.
Questions for authors
  • See critical weakness
1 comment

Thank you for reviewing the paper. Below we state the algorithm and answer your two technical questions.

W1. Presentation.

We agree. The figure text is too small, and no figure shows the algorithm. The revision adds an algorithm box and a problem schematic:

  1. Once, offline: take the first-layer weight W of the frozen action encoder and compute its SVD. The right singular vectors with nonzero singular value are the rows of B (k × d). For the task tier, B is instead the top-k eigenvectors of C = E[∇J ∇Jᵀ], with k at the spectral gap.
  2. In every CEM iteration: sample z in R^k per block, evaluate the cost at a = zB, and refit the mean and variance of z on the elites.
  3. Execute a = μB.

Nothing else in the planner or the model changes. A second variant is simpler still: leave the planner unchanged and project only the executed action onto the rows of B.

W2. Why not simply widen the action embedding layer?

The paper mentioned widening only in one sentence of Section 6, with no experiment. We have now run it: a Cube retrain with the first action layer widened from 10 to 25, matched to one of the paper's retrains (Retrain C, Appendix H) in data, seed, and recipe. Both reach the same validation loss (0.0032 and 0.0031).

150 paired episodeswidened (25)narrow (10)
first-layer rank2510
largest gap in the cost spectrum7.1×, after direction 47.7×, after direction 4
ambient planner68.0%66.0%
sampling the top four cost directions72.7%70.0%

Widening removes the exact blind subspace, as you expected. It does not change ambient planning measurably (+2.0, 95% CI [−2.7, +6.7]), and the cost stays flat along all but about four of the 25 directions. The restriction to those four points the same way on both models (+4.7 [−0.7, +10.0] and +4.0 [−0.7, +8.7]), but neither gain is resolved at this sample size. One training seed per width also limits the conclusion.

Two further points explain why we treat the problem at planning time.

  • The setting is a frozen, released checkpoint. One SVD of its weights takes milliseconds. Retraining takes GPU-days: our 3,000-episode retrain ran for 28 hours on one GPU, and the released recipe uses 10,000 episodes.
  • The planner fault is not limited to the null space. Along exactly blind directions the planner applies no selection at all (Theorem 4.3). Along nearly flat directions the selection is weak, and the planner executes motion it barely selected. On PushT and Two-Room, whose released encoders are already as wide as the block, the cost spectrum also collapses to 2 of 10 directions.

We agree that when retraining is possible, a wider layer is the direct repair for the exact null space. The paper will say this in the main text.

W3. Are the raw actions full rank?

They are full rank. Thank you for asking, because the paper should have shown this. We computed the covariance of all 25-dimensional action blocks in each training set (1.96 million blocks for Cube), standardized per command dimension.

training setdims for 90/95/99% of varianceparticipation ratiomin/max eigenvaluevariance in 15 blind directions
Cube8 / 12 / 216.60.00319.9%
Scene13 / 18 / 238.40.02125.8%
Cube-double17 / 21 / 2410.40.05835.1%

All 25 eigenvalues are nonzero, and covering 95% of the variance takes 12 to 21 dimensions while the bottleneck keeps 10. Your intuition is right in one respect: training orients the kept subspace toward high-variance directions, so the blind subspace has 60% of the dimensions but 20% to 35% of the variance. That is still a fifth to a third of the action variance, and Figure 1 shows that motion along it moves the robot and the object.

The manifold argument also supports the method: even if the actions lay in a low-dimensional subspace, the stock planner would still sample all 25 dimensions with equal variance. We will add the table to the appendix.

W4. Experiments only in simulation.

You are right, and we do not want to understate it. The first item of Section 6 states that all benchmarks are simulated, and we claim nothing about hardware. Within simulation the evidence covers five benchmarks, three retrained Cube models, a three-cube checkpoint, four samplers, and spectra of three TD-MPC2 locomotion models. Three of them involve grasping and placing objects. Simulation lets us pair every comparison by start state and seed and replay each episode deterministically, which the mechanism tests in Section 5.2 require.

We hope these answers make the paper easier to judge. If they address your concerns, we would be grateful if you would reconsider your assessment. We are glad to answer any further question.

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

This paper studies a structural problem in sampling-based planning with latent world models of the joint-embedding family, represented by LeWM. These models stack several consecutive actions into an action block and pass it through a narrow linear layer before it reaches the predictor. When the block is wider than the output of this layer, a null space necessarily exists: the planning cost is exactly constant along these directions, which the authors call the "blind subspace". On the released Cube checkpoint, 15 of the 25 dimensions of the action block lie in the blind subspace.

The authors first show that this problem is real and has a physical effect: perturbations along blind directions cause deviations that accumulate over time in the real environment, while the deviation in the model's imagination stays at the level of floating-point error, so no check based on the model can detect it. They further show that elite selection in the CEM planner exerts no pressure on the blind components, so the executed actions carry an unselected random walk; the authors predict its magnitude theoretically, and the measurements agree with the prediction.

To address this problem, the authors propose a two-tier subspace restriction. The first tier performs one SVD of the frozen action encoder and samples only in the visible subspace, which is proven not to lose the optimum. The second tier estimates the second moment of the cost gradient and, when its eigenvalue spectrum shows a clear spectral gap, samples only along the few directions above the gap. Experiments cover five simulated environments. The first tier improves Cube by 6 percentage points, and a decomposition experiment shows that this gain comes from removing executed noise; it brings no gain on Scene or Cube-double. The second tier improves Cube and Cube-double by 10.7 and 12.3 percentage points, respectively. The authors use dose-response experiments and gradient spectra to explain the differences in gains across environments.

Critical strengths
  1. Novel and meaningful research problem. That action compression makes the planner execute motion it cannot see has received little analysis before. The authors explain its origin in the model architecture and, by contrasting the deviation in the real environment with that in the model's imagination, convincingly show that the problem exists and affects planning, filling a gap in the literature.

  2. Theory and experiments support each other. The proofs of cost invariance and elite decoupling are concise and clear. The predicted magnitude of the executed blind noise matches the measurements inside the solver, and the execution-projection decomposition locates the source of the gain on Cube well.

  3. Thorough experiments with rigorous statistics. The paper uses comparisons paired by start state and seed with exact tests, includes controls such as a random subspace and blind-only sampling, and validates the results across several samplers. The authors honestly report failed predictions and cases where the method does not help, and narrow the scope of their conclusions accordingly.

Critical weaknesses
  1. The proposed method helps only in a narrow range of cases, and the conclusions are not strong. The main value of the paper lies in discovering and analyzing the blind subspace, but the proposed method does not amount to a broadly applicable methodology. The architectural-tier restriction gives a significant gain only on Cube, and none on Scene and Cube-double, which share the same blind subspace. The task tier gives larger gains, but the position of the spectral gap is not stable across retraining, and the number of directions chosen at the gap can be harmful; it also brings no gain on the three-cube checkpoint, whose spectrum is similar to that of Cube-double. The current analysis tools can explain the measured results but cannot yet predict whether the method will help on new models or tasks.

  2. The problem has limited scope. An exact blind subspace appears only when the action block is wider than the first layer of the action encoder. Among the four released LeWM checkpoints, only Cube meets this condition; Scene and Cube-double were trained by the authors with the same configuration. In DINO-WM, the action blocks of all environments are also no wider than the encoder, so no exact blind subspace exists there either. In addition, widening the bottleneck at training time removes the null space directly. I suggest that the authors discuss how common this problem is across world models and action-encoding designs more broadly.

  3. The experiments are somewhat scattered. The main text and appendix contain many experimental branches, controls, and supplementary tests, and different conclusions rely on different environments, budgets, and sampler settings, which makes it hard for readers to tell the main conclusions from the supplementary results.

  4. The writing is hard to follow. The paper is very dense, and a large part of it reports test statistics, paired counts, and preregistration details, so the core ideas and main conclusions do not stand out. I suggest condensing the statistical details in the main text and devoting more space to intuitive explanations of the method and conclusions.

Questions for authors
  1. The project page in the supplementary material states that prior work on action compression in latent world models has reported the effect of post-hoc projection, which this paper explains, but the main text does not seem to cite this work. Could the authors specify which work this refers to and clarify the contribution of this paper relative to it?

  2. The paper finds that the directions above the spectral gap lie almost entirely in the action-repeat subspace, in which the same command is held throughout the action block, and that action-repeat sampling itself also brings a clear improvement. I find this an interesting finding. Can the authors' spectral analysis be used to explain why repeating actions improves planning for this kind of model? A discussion of this point would help clarify the relationship between the effective dimension of the cost and the structure of the actions.

2 comments

Thank you for the close reading, which reached the supplementary project page. Your summary is accurate, and we recognize the weaknesses you list.

Q1. Which prior work does the project page refer to?

It refers to no external work, and we apologize for the confusion. The sentence was left over from an earlier, unpublished draft of our own, written when the measurement study and the planner were separate manuscripts that we merged into this submission. That draft was never posted or submitted. We have removed the sentence.

The effect it mentions is reported in this paper: post-hoc projection onto the top two cost directions lowers PushT success by 8.0 points, and in-loop sampling in the same subspace by 2.0 (Table 8).

Q2. Can the spectral analysis explain why action repeat improves planning?

Yes, and thank you for the suggestion. From the stored spectra, the share of trace(C) inside the action-repeat subspace is 95.0% to 99.2% on the checkpoints with a spectral gap (Cube 97.3%, Cube-double 98.6%) and 89.2% on gapless Scene. That subspace has only 5 of 25 dimensions (2 of 10 on PushT and Two-Room). The cost thus responds almost only to the mean command of a block, and the directions that vary the command inside a block barely move it. The stock planner still samples along them, can barely rank candidates there, and executes the result. Holding the command fixed removes that motion.

The spectrum also says which commands to hold (share of trace(C) per component):

xyzyawgripper
Cube24.1%46.0%17.5%0.06%9.7%
Cube-double21.0%59.3%18.3%0.00%0.06%

This matches Table 8. On Cube-double, holding all five components loses 10.0 points against holding x, y, z. On Cube, where the gripper carries cost mass, it is not worse (+2.3, n.s.). We do not claim the eigenbasis beats the action-repeat basis, which is 3.3 points better on Cube-double. The spectrum adds a diagnosis: it reads the block structure and the components to keep from the frozen model, without knowing the block layout.

One correction: the paper attributes this gain to sampling dilution. A new experiment for Reviewer E9mL shows that projecting only the executed action onto the same directions gives the same gain with the search unchanged (general response, item 1), so the benefit comes from what is executed. We will correct Section 5.3 and add the table above.

W1. Narrow range of benefit; the tools explain but do not predict.

We agree with the second half, and Section 6 says so: no spectral statistic we measured predicts the payoff on a new checkpoint. Two statements hold on every checkpoint we tested.

  • The architectural tier cannot lower the best achievable plan (Proposition 4.2). Its measured change is never significantly negative (+6.0 Cube, +0.0 Cube-double, −3.7 Scene with an interval containing zero).
  • At a deep gap (12× or more; Cube, Cube-double, PushT, Two-Room) the task tier shows no significant loss. Two checkpoints gain 10.7 and 12.3 points, and two show no significant change (−2.0 and +0.0). At shallow gaps (below 8×) the outcomes are mixed, and one cut is harmful (−9.3). We advise the cut only at deep gaps.

Checking a frozen checkpoint is cheap: 2,000 backward passes for the spectrum and four minutes per paired evaluation. The introduction will present the mechanism and the lossless restriction as the main claims, with gains reported per checkpoint.

W2. Exact blindness needs a block wider than the encoder.

Correct, and your count is right: one of four released LeWM checkpoints and none of the eight DINO-WM configurations. Three facts bear on the scope.

  • With the released frame skip of 5 and encoder width 10, any command of three or more dimensions makes the block wider than the encoder. Every OGBench manipulation environment has a 5-dimensional command, and all four manipulation checkpoints in the paper are exactly rank-deficient.
  • The larger gains do not need an exact blind subspace. C collapses on the two full-rank checkpoints too, and the Cube-double gain comes from inside the visible subspace (removing the blind subspace alone gives +0.0).
  • You are right that widening removes the null space directly, and the paper had no experiment on it. We have now run one (general response, item 3). The widened layer is full rank, yet the cost spectrum still collapses to four directions, and ambient planning does not change measurably.

We will add a paragraph on how common the condition is across action-encoding designs.

W3, W4. Scattered experiments and dense writing.

We agree, and we are sorry for the effort this cost you. The revision (general response) adds one main table and an algorithm box, moves test statistics to the appendix, and opens each experiment subsection with the question it answers.

Thank you again. We are glad to answer any further question.

Thank you for the detailed response and the additional experiments. I will keep my original score.

Comments on the paper

1 comment

We thank the three reviewers for the time and care they gave this submission. Several of the criticisms are correct. One question led us to an experiment whose result goes against part of our own explanation, and we report it here in full.

New evidence. All new results use the paper's protocol: comparisons paired by start state and seed, exact McNemar tests, 95% episode-bootstrap intervals. The closed-loop experiments were preregistered before any run.

  1. Execution-only projection on the spectral-gap subspace (Reviewer E9mL). The planner is the unmodified ambient CEM. Only the executed action is projected onto the top-k eigendirections of C. Over 300 paired episodes and six seeds, this gains +8.7 points on Cube (95% CI [+5.3, +12.3]) and +13.3 on Cube-double ([+8.7, +18.3]). In-loop sampling in the same subspaces gains +10.7 and +12.3. The two operators are not distinguishable: the in-loop minus projection difference is +2.0 [−1.7, +5.7] on Cube and −1.0 [−4.0, +2.0] on Cube-double. We had preregistered the prediction that in-loop sampling would be better on Cube-double. That prediction failed.

    A correction that follows. The projection does not change the search, so the Cube-double gain cannot come from the search. We withdraw the paper's attribution of that gain to sampling dilution (abstract, Section 5.3). The data support a different account. The gain comes from removing the component of the executed action along directions where the cost is flat. The measured gains themselves (+10.7, +12.3) do not change.

  2. Effective rank of the raw action data (Reviewer 8MwH). The training actions are full rank. In standardized block coordinates, covering 95% of the action variance takes 12 of 25 dimensions on Cube, 18 on Scene, and 21 on Cube-double. The bottleneck keeps 10. The 15 blind directions carry 20%, 26%, and 35% of the action variance in the three training sets.

  3. Widening the bottleneck at training time (Reviewers 8MwH, kyCw). We retrained the Cube model with the first action layer widened from 10 to 25, matched to a narrow retrain in data, seed, and recipe. The widened layer is full rank, so the exact blind subspace is gone. The cost spectrum still collapses to four directions (largest gap 7.1×, against 7.7× for the narrow model), and ambient planning does not change measurably (68.0% against 66.0%; +2.0, 95% CI [−2.7, +6.7], 150 paired episodes).

  4. Why action repeat helps (Reviewer kyCw). Between 95% and 99% of the trace of C lies in the action-repeat subspace on the checkpoints with a spectral gap (89% on Scene). The cost responds to the mean command of a block and is nearly flat in the variation inside a block.

Planned revision. The tenth page allowed in the discussion phase has room for items a to c.

  • a. One main table with all five checkpoints, both subspaces, and both operators (in-loop sampling and execution projection). In the submission, the Cube-double result appears only in the text of Section 5.3.
  • b. An algorithm box and a schematic of the problem. Larger figure text.
  • c. A paragraph on how common an exact blind subspace is, with the widening experiment and the action-rank analysis.
  • d. The corrected account of the Cube-double gain, with the new projection experiment.
  • e. Paired counts and preregistration details move from the main text to the appendix. Each claim keeps one effect size and one interval.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.