acceptodds
Under review as a conference paper at ICLR 2027

Sparse-TD-JEPA: Improving Zero-Shot Reinforcement Learning With Sparse Latent Task Representations

Abstract

Successor-feature-based zero-shot reinforcement learning methods pre-train a task encoder and successor-feature predictor on offline, reward-free data, then recover a policy at test time by solving a closed-form linear regression over a small labeled dataset. We find that the success rate of this method can be improved by a property that has received little attention: when the task encoder is dense, the predicted reward surface is less concentrated near the goal, producing a weaker gradient for the policy to follow. We address this by introducing a sparsity learning objective into the task encoder that concentrates reward-predictive structure into a smaller effective subspace of the task encoder's latent space. The sparsity penalty is applied during training and requires no change to the closed-form inference procedure at test time. We instantiate this idea as Sparse-TD-JEPA, a direct extension of the recently published TD-JEPA framework, a Temporal Difference (TD) reinforcement framework for the Joint-Embedding Predictive Architecture (JEPA). We evaluate our method across various standard zero-shot performance benchmarks covering various robot manipulation, locomotion, and exploratory navigation tasks. Our method does not only achieve consistent better zero-shot performance than the baseline framework, it achieved the best performance of all established zero-shot baselines tested in most cases.

Primary area:
reinforcement learning
Keywords:
Zero-Shot Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5QSTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Sparse-TD-JEPA adds an l1 penalty to normalized task embeddings, encouraging sparse representations to improve downstream reward inference. It preserves TD-JEPA’s architecture and inference procedure while reporting improved zero-shot RL performance.

Critical strengths
  • Regularizing fixed-norm embeddings encourages sparsity without merely shrinking their magnitude. The method is simple but effective.
  • The results are strong. Average performance improves over TD-JEPA across all 13 domains, covering 65 tasks.
Critical weaknesses

Lack some details, the baselines select the best final-checkpoint configuration separately for each seed, without a stated independent validation set. The sparsity coefficient’s value and search range are unreported, limiting reproducibility and assessment of tuning fairness.

Questions for authors

see weaknesses

1 comment

Dear Reviewer 5QST,

We sincerely thank you for your positive and encouraging review.

We are glad that you found the method simple but effective and noted that the results are strong.

To ensure fair comparison, all baseline in Table 1 are tuned over their own method-specific coefficient, namely the orthonormality coefficient for TD-JEPA, FB, RLDP, BYOL and the Laplacian, and the goal-sampling probability for HILP and ICVF, together with their learning rate.

We tune the sparsity coefficient over [0,0.2][0, 0.2] with a genetic algorithm (Appendix D.9).

To rule out that our gains come from this search, we added Table 10, which reports Sparse TD-JEPA with only a random draw of λsp\lambda_{\mathrm{sp}}, which matches the tuning budget of TD-JEPA.

Domainlr$_\psi$λψ\lambda_\psiTD-JEPAOurs
antmaze-mn10−410^{-4}0.170.4±3.770.4 \pm 3.780.2±2.7\mathbf{80.2 \pm 2.7}
antmaze-ln10−510^{-5}0.157.2±4.357.2 \pm 4.364.0±2.6\mathbf{64.0 \pm 2.6}
antmaze-ms10−410^{-4}0.161.6±4.561.6 \pm 4.573.8±3.1\mathbf{73.8 \pm 3.1}
antmaze-ls10−410^{-4}0.140.6±2.540.6 \pm 2.550.0±2.7\mathbf{50.0 \pm 2.7}
antmaze-me10−410^{-4}120.2±2.420.2 \pm 2.422.0±1.5\mathbf{22.0 \pm 1.5}
scene10−410^{-4}138.4±1.438.4 \pm 1.439.8±1.1\mathbf{39.8 \pm 1.1}
puzzle-3x310−410^{-4}1015.6±1.115.6 \pm 1.118.7±0.7\mathbf{18.7 \pm 0.7}
cube-single10−410^{-4}1034.2±2.934.2 \pm 2.943.0±3.1\mathbf{43.0 \pm 3.1}
cube-double10−510^{-5}103.6±0.83.6 \pm 0.85.3±0.7\mathbf{5.3 \pm 0.7}
walker10−410^{-4}0.1785±7785 \pm 7805±8\mathbf{805 \pm 8}
cheetah10−410^{-4}1689±7689 \pm 7716±5\mathbf{716 \pm 5}
quadruped10−510^{-5}0.1691±5691 \pm 5784±6\mathbf{784 \pm 6}
pointmass10−410^{-4}1479±24479 \pm 24511±17\mathbf{511 \pm 17}

Table 10: Base hyperparameters of Sparse TD-JEPA and zero-shot performance under an equal tuning budget.

Even under this budget, Sparse TD-JEPA outperforms TD-JEPA in all 13 domains.

To make the changes easy to find in the the revised paper, we highlighted all additions in green.

We would be grateful if you would consider these clarifications in your final assessment, and we are happy to answer any further questions.

We thank you again for your careful reading and your constructive comments!

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

The paper adds an ℓ1 penalty to normalized task embeddings in TD-JEPA to improve zero-shot reinforcement learning. Across 65 tasks, it reports higher domain-level means than TD-JEPA and associates these improvements with more informative inferred reward landscapes.

Critical strengths

The method is simple and introduces no additional inference-time computation, making it a potentially useful improvement to existing systems. Reported improvements span all 13 evaluated domains and multiple training seeds, providing encouraging evidence beyond isolated tasks.

Critical weaknesses
  1. Limited conceptual novelty. The methodological change is a standard ℓ1 penalty on normalized task embeddings, while TD-JEPA’s architecture, predictive objectives, policy learning, and inference remain unchanged. The paper itself discusses prior work showing that sparse activations reduce interference in TD-based value learning. Its contribution therefore requires a distinctive insight into zero-shot reward inference beyond applying this familiar regularizer to another encoder. The benchmark improvements demonstrate practical usefulness, but the paper does not sufficiently distinguish its proposed mechanism from established benefits of sparse representations. Neither a new inference principle nor a convincing explanation specific to zero-shot control is established.

  2. Unsupported core premise. The motivation treats dense feature coordinates as a cause of spatially diffuse inferred rewards, although these are different properties. An orthogonal rotation can turn sparse embeddings into dense ones while preserving the least-squares reward predictions, including the spherical projection used in inference. Coordinate density alone therefore cannot explain reward-landscape sharpness. The paper needs to identify the additional geometric property changed by training and establish its connection to reward inference, rather than treating sparsity itself as the explanation.

  3. Incomplete theoretical support. The theoretical discussion characterizes the numerical sparsity of individual embeddings, which is not equivalent to the rank of the dataset feature matrix or the dimension of a shared reward-relevant subspace. Different sparse embeddings can collectively span the full latent space, and the paper acknowledges that shared support is not enforced. Consequently, reducing per-example effective dimension does not establish lower regression variance, smaller approximation error, or improved policy performance. The connection between the stated theoretical properties and the claimed control benefits remains incomplete.

  4. Weak mechanistic evidence. For goal-conditioned tasks, the analysis measures linear correlation between inferred immediate rewards and Euclidean distance to the goal. This does not directly measure reward-reconstruction accuracy or control-relevant progress, particularly in environments with obstacles or complex object interactions. A sharper immediate-reward peak also does not automatically provide better guidance to the pretrained policy. The reported correlations therefore offer descriptive evidence, but do not establish that improved reward concentration causes the performance gains or explain how the proposed mechanism operates across the evaluated task families.

  5. Missing targeted controls. Although the experiments include numerous established algorithms, they do not adequately test the proposed explanation. The stated problem concerns nuisance coordinates and spurious correlations in unregularized least squares, yet regularized reward inference, reduced-dimensional embeddings, and alternative sparsification strategies are not compared. The experiments also do not directly connect the penalty strength, realized sparsity, shared feature structure, and downstream return. Such controls are essential to establish the specific value of the proposed intervention and distinguish its claimed mechanism from broader regularization or optimization effects.

Questions for authors

See the weaknesses.

9 comments

Dear Reviewer tXwP,

We identify an underexplored limitation of successor-feature zero-shot reinforcement learning (RL): the geometry of the task encoder controls how sharp the inferred reward is. We address it with a one-line fix and a complete theoretical account. Under normalized task embeddings, as in TD-JEPA, the standard regularizers are provably inert, and ℓ1\ell_1 becomes a pure effective-dimension pressure. Prior work studies activation sparsity in value networks for online control. Our analysis concerns zero-shot reward inference and is new.

Your rotation argument is formally covered by our Appendix A.7. It leads directly to our explanation: the penalty changes the learned feature span, and the NN-independent gap in Figure 5 is the evidence. New controls further show that the gains come from the sparsity penalty itself, separate from orthonormality pressure.

To answer your questions on realized sparsity and broader regularization effects, we ran two new studies during the rebuttal period. Both are marked in green in the revised paper.

New experiments added for your review

1. Direct geometric measurement (new Appendix A.8, Table 2). You asked us to connect the penalty to realized sparsity and to separate its effect from broader regularization. We therefore measured the per-example effective dimension and the orthonormality deviation of trained encoders. Each comparison uses paired runs that share the seed and all base hyperparameters and differ only in λsp\lambda_{sp}. Each quantity is computed on the same 10,000 dataset states.

2. Large orthonormality sweep (new Appendix E, Table 11). You asked whether broader regularization effects could explain the gains. We therefore retrained TD-JEPA over a wider and finer grid of the orthonormality coefficient λψ\lambda_{\psi}, the TD-JEPA hyperparameter closest to encoder geometry. The sweep covers 37 configurations and 185 newly trained models (548.7 GPU-hours) across five domains, each with five seeds.

We hope these results address your concerns, and we would be grateful if you would consider increasing your score in light of them. We address each point in detail below.

Conceptual novelty

Our contribution addresses a central step of successor-feature zero-shot RL: inferring a reward from a few labeled states by least squares on the task encoder. We identify that the geometry of the task encoder limits how sharp this inferred reward is, and we give a one-line fix with a complete theoretical account. The fix is a training-time loss on the task embedding, so it can be added to any method that infers rewards this way. Our analysis shows that this setting differs from the familiar one in three ways, and Appendix A proves each of them.

1. The standard regularizers are provably inert here. TD-JEPA projects task embeddings onto a sphere of radius dψ\sqrt{d_\psi}. Theorem 1 shows that every rotation-invariant regularizer on the projected embedding has zero gradient. This covers ℓ2\ell_2, squared ℓ2\ell_2, and isotropically smoothed norms (Corollary 2). Training with any of them follows exactly the same parameter trajectory as TD-JEPA. Weight decay on the output layer only moves parameters along function-preserving rays (Remark 2). The familiar toolbox is therefore inert in this architecture.

2. The ℓ1\ell_1 penalty becomes a pure effective-dimension pressure. On the sphere, ∣x∣1=dψ,deff(x)\\|x\\|_1 = \sqrt{d_\psi \\, d_{\mathrm{eff}}(x)} holds exactly (Proposition 1). The norm is fixed, so the penalty acts on direction alone (Corollary 1).

3. The penalty fixes a quantity that TD-JEPA leaves undetermined. The base objective is exactly invariant to rotations of the task space (Proposition 6). Per-example effective dimension varies from 11 to Θ(d)\Theta(d) within a single orbit of equivalent solutions (Lemma 2, Remark 7). Neither the predictive losses nor the orthonormality loss select a value. The ℓ1\ell_1 term is the minimal coordinate-anisotropic addition that does.

New in the rebuttal. We now verify point 3 directly on trained encoders (Appendix A.8, Table 2). The penalty lowers deffd_{\mathrm{eff}} in 9 of 9 domains. The orthonormality deviation ∣Σ−I∣F\\|\Sigma - I\\|_F stays at the level of the base method. The penalty and the orthonormality loss therefore act on different quantities, exactly as Lemma 2 predicts.

The practical value is also clear. The change is one line of code with zero inference-time cost. It gives a higher mean than TD-JEPA in 13 of 13 domains and the best mean against eight established baselines in 9 of 13.

A simple change with a precise theoretical account and consistent gains is, in our view, a strong contribution.

Core premise

Our analysis addresses your rotation argument formally. Appendix A.7 proves it in full generality. Proposition 6 shows that zero-shot inference is equivariant under orthogonal rotations. Lemma 1 shows that the spherical projection preserves every scale-invariant readout. Remark 7 states the consequence explicitly: re-expressing a fixed set of features in a sparser basis leaves every inferred landscape unchanged.

Our explanation rests on the span of the learned features on the labeled states. Proposition 5 connects this span to reward inference exactly:

E[1N∣V^−r∣22]=1N∣(I−PΨ),r∣22+σ2,rank(Ψ)N.\mathbb{E}\Big[\tfrac{1}{N}\\|\hat V - r\\|_2^2\Big] = \tfrac{1}{N}\\|(I - P_\Psi)\\, r\\|_2^2 + \frac{\sigma^2 \\, \mathrm{rank}(\Psi)}{N}.

The first term depends only on the feature span. Training with the penalty changes this span. The penalty breaks the rotational symmetry during optimization, and the encoder learns different feature functions (Propositions 3 and 4, Remark 7).

Corollary 3 provides the evidence that the span changed. A gap caused by estimation variance decays as 1/N1/N. A gap caused by the span is constant in NN. In Figure 5 the sparse-versus-dense gap stays approximately constant from N=104N = 10^4 up to the full dataset. This is precisely the signature of a change in the feature span.

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

The authors propose adding L1 regularization to the task latent encoder of TD-JEPA, a zero-shot RL method based on Successor Features and self-predictive losses. The proposed change can be implemented in a single line of code and leads to a surprisingly big performance gain across standard zero-shot RL benchmarks. The authors provide extensive empirical validation of their approach across two statistical tests, showing that the sparsity they introduce leads to better reward inference across different inference budgets, as well as, an extensive appendix of theoretical results showing that L1 regularization is a very good choice for a convex regularization that increases the numerical sparsity of features that live on a hypersphere.

Critical strengths

The paper identifies a good improvement to a state-of-the-art unsupervised RL method that is easy to implement. The authors provide ample evidence that their method improves over prior work across a wide range of tasks and environments, as well as providing theoretical justification for their approach.

Critical weaknesses

In its current form the paper has one critical weakness: It is unclear whether the L1 regularization has an effect on the orthogonality of the feature covariance $ \Sigma_\rho = E_{s \sim \rho} [\psi(s) \psi(s)^\top ] $ which can have a large impact on performance in successor feature methods. If it does, the added regularization might have the same effect as increasing λψ\lambda_\psi which would mean that the TD-JEPA baseline could be tuned to achieve the same performance. The authors provide a theoretical analysis that their regularization induces sparsity and I agree with their intuition that this should improve successor feature learning. But L1 regularization might also have other effects, crucially forcing the feature covariance closer to orthogonal, and they do not control for this effect. For this reason I would not recommend accepting this paper until more evidence is provided.

update: these concerns have been addressed.

Questions for authors
  1. For acceptance, I would need to see evidence that L1 regularization offers more than just additional orthogonality regularization. Can you show that the TD-JEPA baseline does not pick the highest λψ\lambda_\psi in a wide and granular enough grid search? The current grid search covers 2 (OGBench) and 3 (DMC) values respectively.
  2. Can you investigate how close to orthogonal Σρ\Sigma_\rho is for the sparse and non-sparse versions of TD-JEPA? does L1 increase orthogonality?
  3. If I am not mistaken, the actual choice of λsp\lambda_\text{sp}, or at least its search-grid, is not specified anywhere in the paper. Could you add it?
7 comments

Dear Reviewer aWcY,

Your insightful review has been highly appreciated and was very useful to us. We are glad you agree with our intuition that sparsity should improve successor-feature learning, and we thank you for the time you spent on our paper!

Furthermore, we apologize for the somewhat late reply. We wanted to address all your points with new experiments, not arguments alone. This required training 185 new models over 548.7 GPU-hours. Now that we have the results, we would like to complete our case.

You stated that for acceptance you would need evidence that ℓ1\ell_1 regularization offers more than additional orthogonality regularization. Below we provide three independent pieces of evidence, each addressing one of the concerns you raised. Furthermore, we uploaded a revised PDF that extends the paper with these new results. To make the changes easy to find, all new additions in the revised paper are highlighted in green.

Compact summary of the findings: Tuning TD-JEPA's orthonormality over a wider and finer grid, with twice our tuning budget, does not reproduce the gains of Sparse TD-JEPA (Q1). Measuring the covariance directly shows that the penalty does not make it more orthogonal (Q2). Equalizing the tuning budget preserves the improvement in all 13 domains (Q3). All new results and the full experimental details are included in the revised paper (Appendices~A.8, D.9 and E).

We hope this answers your question. If you are convinced with our arguments and our extended experiments findings, we would highly appreciate an increased score.

We thank you very much again for your insightful and constructive review!

Q1: Wider grid for TD-JEPA.

If the ℓ1\ell_1 penalty acted as additional orthogonality regularization, two things should happen when TD-JEPA is tuned over a wider grid of λψ\lambda_\psi. Its best λψ\lambda_\psi should move toward stronger orthonormality, and the tuned TD-JEPA should close the gap to Sparse TD-JEPA. We tested this directly. We extended the original grid of Table 9 to a half-decade logarithmic grid of 7–8 values per domain, spanning three or more orders of magnitude, on five domains with five seeds each. This amounts to 185 newly trained models and 548.7 GPU-hours (Table 11).

Domain grouplr$_\psi$λψ\lambda_\psi
OGBench Antmaze{10−4,10−5}\{10^{-4},10^{-5}\}{0.1, 1}\{0.1,\,1\}
OGBench Manipulation{10−4,10−5}\{10^{-4},10^{-5}\}{1, 10}\{1,\,10\}
DMC{10−4,10−5}\{10^{-4},10^{-5}\}{0.01, 0.1, 1}\{0.01,\,0.1,\,1\}

Table 9: Original TD-JEPA hyperparameter grid.

λψ\lambda_\psiquadrupedcheetahantmaze-mnantmaze-mscube-single
0.003728±10728 \pm 10595±27595 \pm 27–––
0.01743±21743 \pm 21642±20642 \pm 2086.0±3.986.0 \pm 3.967.6±4.867.6 \pm 4.8–
0.03703±23703 \pm 23655±31655 \pm 3183.6±3.383.6 \pm 3.362.4±3.362.4 \pm 3.3–
0.1721±34721 \pm 34646±38646 \pm 3876.4±5.376.4 \pm 5.365.6±4.565.6 \pm 4.514.8±1.414.8 \pm 1.4
0.3702±29702 \pm 29655±39655 \pm 3976.8±4.676.8 \pm 4.666.8±3.166.8 \pm 3.124.0±6.024.0 \pm 6.0
1682±15682 \pm 15665±13665 \pm 1370.8±3.370.8 \pm 3.352.8±4.852.8 \pm 4.826.0±2.826.0 \pm 2.8
3518±37518 \pm 37641±33641 \pm 3374.8±4.474.8 \pm 4.453.6±5.253.6 \pm 5.230.0±1.730.0 \pm 1.7
10479±35479 \pm 35669±8669 \pm 876.4±2.076.4 \pm 2.051.6±6.051.6 \pm 6.028.4±1.928.4 \pm 1.9
30––––29.6±7.929.6 \pm 7.9
100––––33.6±5.333.6 \pm 5.3
TD-JEPA, original grid74366576.465.628.4
TD-JEPA, wider grid74366986.067.633.6
Sparse TD-JEPA784±6784 \pm 6716±5716 \pm 583.2±4.883.2 \pm 4.873.6±6.173.6 \pm 6.143.0±3.143.0 \pm 3.1

Table 11: TD-JEPA over a wider grid of λψ\lambda_\psi. Mean and standard error over five seeds of the final-checkpoint score (return on DMC, success rate in % on OGBench). A dash marks values outside the grid of that domain.

Neither prediction holds.

  • The gap persists. With 7–8 configurations per domain, TD-JEPA receives roughly twice the tuning budget of Sparse TD-JEPA, and it still stays below Sparse TD-JEPA in 4 of 5 domains, by 41 on quadruped, 47 on cheetah, 6.0 on antmaze-ms and 9.4 on cube-single. On antmaze-mn the two are within one standard error.
  • TD-JEPA moves toward weaker orthonormality. On quadruped and both antmaze domains, TD-JEPA performs best at λψ=0.01\lambda_\psi = 0.01, the weakest orthonormality on its grid. If ℓ1\ell_1 helped by adding orthonormality, TD-JEPA would improve with larger λψ\lambda_\psi, and on these domains it improves with smaller λψ\lambda_\psi.
  • More orthonormality is not enough. On cheetah and cube-single, where TD-JEPA does prefer stronger orthonormality, it receives up to λψ=10\lambda_\psi = 10 and 100100, and still falls 47 and 9.4 points short of Sparse TD-JEPA.

Note on hardware. Tables 10 and 11 were run on different hardware. Configurations that appear in both tables agree within two standard errors.

Q2: Does ℓ1\ell_1 make Σ\Sigma more orthogonal?

Our submission had already addressed this question theoretically. Lemma 2 proves that the orthonormality loss is blind to per-example effective dimension: two encoders can have identical second moments Σ=E[ψψ⊤]\Sigma = \mathbb{E}[\psi\psi^\top], and therefore identical orthonormality loss, while their effective dimensions differ by a factor of order dd. The ℓ1\ell_1 penalty acts on exactly this quantity, so the theory predicts that it sparsifies the embeddings without changing how close Σ\Sigma is to the identity.

Following your suggestion, we tested this prediction empirically. Table 2 compares trained encoders that share seed and base hyperparameters and differ only in the sparsity penalty, in nine domains. The ℓ1\ell_1 penalty lowers the per-example effective dimension in all nine domains, while it brings Σ\Sigma closer to orthogonal in only three of them, and by at most 0.0330.033. The ℓ1\ell_1 penalty therefore changes the property it was designed to change and leaves orthogonality essentially where TD-JEPA puts it, exactly as Lemma 2 predicts.

Domain∥Σ−I∥F/d\Vert\Sigma - I\Vert_F/\sqrt{d} TD-JEPA∥Σ−I∥F/d\Vert\Sigma - I\Vert_F/\sqrt{d} Oursdeffd_{\mathrm{eff}} TD-JEPAdeffd_{\mathrm{eff}} Ours
antmaze-mn1.5591.57828.9125.30
antmaze-ln1.6231.65231.1229.18
antmaze-ms1.5881.60029.4027.57
antmaze-ls1.8421.80928.4725.65
scene0.6390.67532.1831.91
puzzle-3x30.0850.09132.0731.86
walker1.2851.32730.4429.50
cheetah0.4530.45232.2132.13
quadruped0.9130.90831.8531.67
Domains where Ours is lower3/99/9

Table 2: Orthonormality deviation and per-example effective dimension of the task encoder for paired runs that share seed and base hyperparameters, averaged within each domain. Lower ∥Σ−I∥F\|\Sigma - I\|_F means second moments closer to isotropy.

Comments on the paper

1 comment

General Response to the Chairs and All Reviewers

Dear Program Chairs, Senior Area Chairs, Area Chairs and Reviewers,

We thank all three reviewers for their time and careful reading. Your comments helped us strengthen the paper. We answer all points not with arguments alone, but with new extended experiments and evidence. All additions in the revised PDF are highlighted in green.

Strengths recognized by the reviewers

  • Strong and consistent gains. Reviewer aWcY notes "a surprisingly big performance gain," and Reviewer 5QST states that "the results are strong." Sparse TD-JEPA improves over TD-JEPA in all 13 domains and 65 tasks (aWcY, tXwP, 5QST).
  • Sound. Reviewer 5QST reports no substantive soundness concerns. Reviewer aWcY agrees with our intuition that sparsity should improve successor-feature learning.
  • Extensive empirical validation. Reviewer aWcY highlights validation with two statistical tests across different inference budgets.
  • Theoretical justification. Reviewer aWcY recognizes the theoretical results showing that ℓ1\ell_1 is a well-suited convex regularizer for features on a hypersphere. Reviewer 5QST notes that regularizing fixed-norm embeddings induces sparsity without shrinking magnitude.
  • Simple and practical. Reviewer 5QST calls the method "simple but effective." It is a one-line change with zero inference-time cost (aWcY, tXwP, 5QST).

Questions raised and where we answered them

Wider grid for the TD-JEPA orthonormality coefficient λψ\lambda_\psi

We supplied a new experiment in Q1 for Reviewer aWcY (new Appendix E, Table 11). We swept λψ\lambda_\psi over 37 configurations and 185 models (over 548.7 GPU-hours). TD-JEPA received twice our tuning budget and still stays below Sparse TD-JEPA in 4 of 5 domains.

Does ℓ1\ell_1 make the feature covariance more orthogonal?

We supplied a new measurement in Q2 for Reviewer aWcY (new Appendix A.8, Table 2). The penalty lowers the per-example effective dimension in 9 of 9 domains. It brings the covariance closer to orthogonal in only 3 of 9 domains, as Lemma 2 predicts.

Value and search range of the sparsity coefficient λsp\lambda_{sp}

We answered this in Q3 for Reviewer aWcY and in our reply to Reviewer 5QST (Appendix D.9, Table 10).

Tuning

We answered this in our reply to Reviewer 5QST (Table 10). With a single random draw of λsp\lambda_{sp}, which matches the tuning budget of TD-JEPA, Sparse TD-JEPA still outperforms TD-JEPA in all 13 domains.

Link between theory and control performance

We answered this in "Theoretical support and control benefits" for Reviewer tXwP. We lay out four links from the penalty to control. Each link is backed by a proof or a direct measurement.

Targeted controls and alternative explanations

We answered this in "Targeted controls" for Reviewer tXwP (Tables 2, 10, 11 and Appendix A.9). These controls exclude tuning and orthonormality as alternative causes. Appendix A also rules out rotation-invariant penalties, dense ℓp\ell_p penalties and dropout analytically.

On the assessment of Reviewer tXwP

Two reviewers find the results significant and the effects clear. Reviewer aWcY describes the gain as "surprisingly big" and the validation as "extensive." Reviewer aWcY's only critical weakness was the orthogonality question, which we have now answered with direct experiments. Reviewer 5QST finds the results strong and reports no soundness concerns. Reviewer tXwP also acknowledges that the improvements "span all 13 evaluated domains and multiple training seeds."

Reviewer tXwP's "Questions For Authors" field only refers back to the weaknesses. Most of these weaknesses state that a link "is not established" or "remains incomplete." They do not say which result would change the assessment. We respectfully note that this makes the concerns hard to resolve. We have nonetheless answered every point with proofs from the paper and new experiments.

We thank all reviewers again and are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.