acceptodds
Under review as a conference paper at ICLR 2027

Representation Selection in Capacity-Limited LeJEPA

Abstract

Self-supervised Joint-Embedding Predictive Architectures (JEPAs) learn from correlated observations of the same system. Existing LeJEPA theory shows that, when the representation dimension matches the number of latent variables and all latent variables evolve at the same rate, the training objective can linearly recover the latent variables. In practical models, however, the representation dimension is often smaller, and there is still no complete theory of how limited representation capacity is allocated. We study latent variable selection in LeJEPA under limited representation capacity and analyze how its distributional constraint affects this selection. We show that, when latent variables evolve at different rates, the model ranks both the latent variables and nonlinear functions of them. As a result, functions of higher order of slower latent variables can replace faster latent variables in the learned representation. LeJEPA's distributional constraint can rule out simple polynomial solutions of higher order, but it can still allow multiple representation coordinates to depend on the same slow latent variable. This source reuse can leave predictable information about the next representation outside the selected span. Based on this analysis, we introduce an additional conditional mean constraint to reduce this information loss. Under the Gaussian model, the two constraints yield linear invariant structure, and alignment selects the leading linear subspace when the cutoff is strict. Our experiments support the predicted selection boundary and the effect of the proposed correction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, LeJEPA, representation learning, latent variable recovery

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

This paper studies representation selection in capacity-limited LeJEPA and shows that Gaussianity still permits latent source reuse. It introduces a transition-closure constraint and establishes exact linear recovery and stable near linearity under the stated assumptions, supported by controlled experiments.

Critical strengths

The paper addresses a meaningful question and proposes a well-motivated transition-closure constraint. The theoretical results on exact linear recovery and stable near linearity are valuable, and the experiments provide targeted evidence for the mechanisms identified by the theory.

Critical weaknesses

The connection between the theoretical conditions and the implemented training procedure is not sufficiently validated. The recovery guarantees require control of Gaussianity error and transition-closure error, together with a uniform predictability floor. However, the implementation uses soft Gaussian regularization and a finite critic, whose small objective does not necessarily imply a small true closure defect. Although the experiments demonstrate recovery in some settings, they do not sufficiently examine its sensitivity to critic capacity and optimization budget or explain the settings where recovery fails. Consequently, the reliability of the proposed correction remains insufficiently supported.

If the authors can adequately address these validation concerns in the rebuttal, I would consider raising my score to 3.

By the way, the fonts in some figures are excessively large. Reducing their size would improve the presentation. This is only feedback from reading the paper and does not affect my score.

Questions for authors
  1. Which regimes of the original Gaussian-constrained objective are fully characterized, and what remains open under unequal transition rates?
  2. The synthetic CSV contains negative closure_defect values despite the nonnegative clamp in zcm_rf, as well as fields absent from the current output code. Which code version and post-processing procedure generated these results?
5 comments
AuthorsOpenReview ↗

Empirical Validation of Theoretical Conditions (W1)

W1. Connecting Theoretical Conditions to the Training Procedure

Thank you for the suggestion. We have added experiments for independent measurement of the theoretical conditions, comparisons of critic capacity and optimization budget, and analysis of failure cases.

  1. With a 4-dimensional latent variable and a 2-dimensional representation, we examined the following four quantities:
MethodTarget recovery error ↓Independent closure error ε\varepsilon ↓Fourth-moment discrepancy Δ4\Delta_4 ↓Predictability floor bb
Gaussian2.41960.025850.50830.8728
ZCM0.01270.001600.54380.2043

In this longer training run, both the target recovery error and the independently measured closure error improve, while Δ4\Delta_4 increases slightly.

  1. Critic capacity and optimization budget also matter. We varied the critic used during training and measured target recovery error and independent closure error:
Critic width / updates per stepTarget recovery error ↓Independent closure error ε\varepsilon ↓
32 / 30.940220.01268
128 / 30.060500.00588
512 / 30.060690.00596
128 / 11.787090.01614
128 / 100.050510.00230

These results show that an undersized critic or insufficient updates can impair recovery. The default configuration, with width 128 and three updates per step, already performs well. Further increasing the capacity or the number of updates brings limited gains. The independent closure error improves alongside the recovery error, indicating that the effect is not merely a reduction in the score reported by the training critic.

The results suggest that a critic that is too small may underestimate the closure defect. Once its expressive capacity is adequate, however, increasing it further brings no clear benefit. This indicates that a suitably configured critic has a sufficiently small approximation gap to serve as an optimization surrogate during training.

  1. We observe two situations that can affect performance: insufficient critic capacity or optimization, and the inability of the soft constraint to maintain a good representation throughout prolonged training. The former can be addressed by appropriately increasing the capacity and optimization budget, while excessively large settings do not bring further gains.

The goal of this paper is to recover the linear latent space, rather than to optimize task relevance, so task relevance is not our focus here. Future work could combine task-relevance methods with task-specific information and modify alignment accordingly, allowing it to select latent variables that better match the task.

These experiments address the original manuscript’s lack of direct condition checks and sensitivity evidence, while clarifying both the issues that have been improved and the remaining limitations.

AuthorsOpenReview ↗

Scope of the Gaussian-Constrained Objective and Code Provenance (Q1–Q2)

Q1. Characterization of the Gaussian-Constrained Objective

We organize the regimes in which the Gaussian-constrained objective is fully characterized into three levels.

  1. For the whitening-only problem, which requires only zero mean and identity covariance, we have fully characterized top-m selection through the complete Hermite spectrum. The selected modes may be the latent variables themselves or the squares and higher-order functions of the same slow latent variable. This level is fully characterized, including cases with tied eigenvalues.

  2. Under the exact Gaussian constraint, we have a complete characterization for equal latent transition rates, representation dimension m=1m = 1, and a heterogeneous-rate regime in which the linear factors are sufficiently dominant. Ordering the rates as ρ1≥⋯≥ρd\rho_1 \ge \cdots \ge \rho_d, this regime satisfies

ρm≥max⁡(ρm+1,ρ12).\rho_m \ge \max(\rho_{m+1}, \rho_1^2).

In other words, the weakest retained linear factor is at least as predictable as both any omitted linear factor and the strongest competing second- or higher-order mode. In these cases, any Gaussian representation attaining the whitening spectral optimum must be linear, and the remaining ambiguity comes only from ties among linear transition rates.

The problem becomes substantially harder when ρm\rho_m is smaller than higher-order modes of a slow latent variable. Completing this broader characterization would require establishing global optimality for arbitrary measurable Gaussian representations or constructing the full family of optimizers. The existing source-reuse construction provides a counterexample showing that a leading linear representation is not optimal in some cases, but it does not fully characterize the optimizer family or provide an exact decomposition throughout this entire difficult regime.

  1. When Gaussianity and exact transition closure hold simultaneously, we characterize the feasible representations as linear RR-invariant subspaces. Under a strict spectral cutoff, alignment further selects the leading linear subspace. This is the central result of our proposed correction and does not depend on first solving for all optimal representations without closure in the general unresolved regime.

We will clearly distinguish the regimes that have been resolved from those that remain open, making the scope of the theoretical results more explicit.

Q2. CSV Fields and Code-Version Consistency

Thank you for checking this carefully. After tracing the discrepancy, we confirmed that it arose from a version mismatch introduced during code cleanup.

The closure_defect field in the original CSV corresponds to zcm_eps_norm in the original results. It is a normalized, signed latent-Hermite held-out dual and can therefore take negative values. During cleanup of the submitted code for anonymous review, the same-named field was connected instead to the observation random-feature diagnostic after nonnegative clamping.

Our item-by-item check confirms that all results reported in the paper are consistent with the original experimental records. We will carefully correct the code to match the original experimental implementation.

Thank you for the additional experiments. They substantially address my concern about the absence of critic-capacity and optimization-budget studies. My original review raised the question of how sensitive recovery is to the critic’s optimization budget. The additional results in your response to Reviewer PZ8g directly bear on this concern. At the same critic width, increasing the number of updates per step from 10 to 20 or 40 substantially worsens both target recovery and the independently estimated closure error. Could you explain this deterioration and clarify how the update budget should be chosen in practice?

On a separate note, I was disappointed that the authors did not respond at all to my earlier comment about the oversized fonts in some figures, even though this issue still does not affect my rating.

Reviewer_rL7qOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper tries to solve in problem of encoder concentrating on slow moving features [1] in JEPA based frameworks. The paper comes up closure method that tries to recover fast moving features/latent variables. It has a linear predictor and an auxiliary critic. The critic adds significant overheard to the method. The paper shows that this setup and new objective learns the predictive features in couple of toy environments.

[1] Sobal et al. Joint embedding predictive architectures focus on slow features. NeurIPS SSL Theory and Practice Workshop 2022. https://arxiv.org/abs/2211.10831

Critical strengths

Strengths:

  • The paper starts with why LeJEPA objective (SIGREG) of representations doesn't guarantee all the latent variables getting captured. They justify it by theoretrical proofs (Theorem 1 an 3).
  • The paper shows that sigreg with closure method is able to capture various latents in various synthetic environments. They also do nice job of doing various controls and ablations for the setup.
Critical weaknesses
  • Environment Setups : The paper does evaluation in very limited settings: a environment with nonlinear mixtures of Gaussian latent variables, and a rendered disk environment with four different state factors. While these environments are good for diagnosing various factors, There are very limited for real visual learning problems.
  • Non Linear temporal dynamics : While the closure ensure that linear Gaussian latent dynamics and observations that preserve the underlying state, It may not help/ hurt the non linear features learning of encoder. It could potetially represent collision-sensitive motions less strongly and give more capacity to easier features like velocities.
  • The latent dimensions are too small: The underlying state dimension and the embedding size (2 dimensions) need be increased to see whether the findings persist beyond this narrow setting.
Questions for authors

Overall, The above method doesn't solves any real problem like LeWM [2] or PLDM [3] papers. These works have mechanisms to preserve linear features implicitly by design or explicit loss terms. The paper doesn't actually do eval of techniques in more real environments where non-linear features may also play important role. The current environments are small and limited. I think, These calrifications can add further value to the paper :

  • It would be interesting, if authors can show the benefits of these representations in more realistic setting like PushT, PointMaze environments. The emphasis is on environments with slow moving or static distractors.

  • Can the experiments jointly increase the number of underlying factors and the embedding dimension, reporting recovery quality of features?

  • Please cite and discuss [1–3]. Please present analysis extends earlier observations of slow latents in [1], and how closure relates to the prediction, inverse-dynamics, and representation smoothness objectives used in PLDM and LeWM.

[1] Sobal et al. Joint embedding predictive architectures focus on slow features. NeurIPS SSL Theory and Practice Workshop 2022. https://arxiv.org/abs/2211.10831

[2] Sobal et al. Learning from reward-free offline data: A case for planning with latent dynamics models. 2025. https://arxiv.org/abs/2502.14819

[3] Maes et al. LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels. 2026. https://arxiv.org/abs/2603.19312

2 comments
AuthorsOpenReview ↗

Structural Recovery, Task Relevance, and More Realistic Evaluation (W, Q1)

W. Structural Recovery versus Task-Relevant Information

Thank you for raising this point. We will clarify the paper’s objective. We distinguish whether the representation space has an appropriate structure from whether its retained information suits a particular task. We address the former; the latter requires additional task information. The two are complementary.

  1. Linear recovery does not require the encoder or task readout to be linear. The term “linear” in our theorem refers to the latent variable Z: h(Z)=AZh(Z) = AZ, a linear invariant subspace of the latent state space. If X=g(Z)X = g(Z) is an invertible nonlinear mapping, f(X)=Ag−1(X)f(X) = Ag^{-1}(X) can still be highly nonlinear. Nonlinear readouts can compute collision events or rewards from retained state information. Our concern is that limited capacity allows multiple coordinates to reuse one slow source through different nonlinear functions, leaving other latent directions unrepresented. Under the theorem’s assumptions, Gaussianity and closure rule out this ambiguity and recover linear coordinates of the selected latent subspace; the underlying motion dynamics need not be linear. We provide an additional collision experiment in Q1.

  2. Correct structure does not imply task optimality, and our method does not guarantee retention of all task-relevant information. Among representations satisfying the structural constraints, alignment still ranks directions primarily by predictability. For example, with only two representation dimensions, an irrelevant background with temporal correlation 0.99 may be selected alongside a task factor with correlation 0.8, while another task factor with correlation 0.7 is omitted. This can happen even if all representations are linear. Nor can the method guarantee that collision-sensitive factors receive more weight than easier-to-predict factors. This is a limitation of the original method’s objective and scope.

We distinguish two complementary levels. Structural identification asks how a limited number of representation dimensions are organized and used. Once the subspace is identified, task-aware selection asks which information is most valuable for the current task. Task supervision, rewards, goal conditioning, or nuisance invariance can provide additional information for the latter.

When m<dm < d, a task-relevant factor discarded at the first stage cannot be recovered by a downstream readout. This motivates studying representation structure alongside task-aware selection, since a downstream readout cannot recover task-relevant information that the representation has discarded. Our work provides a structural basis for subsequent processing. Alignment could be modified using task relevance to obtain a task-constrained selection method, extending this line of work to a broader setting.

  1. The PointMaze experiment also trains with collision-prediction and nuisance-invariance losses, rather than merely evaluating position or collision readability after training. Under matched action, task, and invariance conditions, AP is 0.4230 with ZCM and 0.3967 without it. This supports combining structural and task constraints, with an additional separable gain from ZCM. Our work provides structural-identification results under a specified model; the new experiment tests their compatibility with task constraints.

Q1. Evaluation in More Realistic Visual and Real-Data Settings

Thank you for the suggestion. The original manuscript used controllable Gaussian latent states and a four-factor disk environment to validate the theoretical mechanism. To examine more complex visual settings, we added a more complex PointMaze visual simulation and experiments on two real-world datasets.

PointMaze includes maze walls, position, velocity, collisions, and a goal marker. Adding Closure increases average position from 0.2291 to 0.4092, but collision-prediction AP only increases from 0.1024 to 0.1061. After adding action, collision-task, and nuisance-invariance constraints, collision AP increases from 0.0842/0.0879 with unconditional ZCM to 0.4200/0.4230. The corresponding expanded model without ZCM obtains 0.3979/0.3967, showing that ZCM still provides an additional gain.

For real-world data, SMD records 38 server operating metrics. After compressing them to 8 dimensions for anomaly detection, AP increases from 0.1946 to 0.3238. On the ETTm1 power-transformer dataset, forecasting NMSE decreases from 1.0202 to 0.9444. These results show that ZCM can improve some tasks on real-world records. The additional experiments provide evidence for the theoretical mechanism in more realistic settings.

AuthorsOpenReview ↗

Higher-Dimensional Evaluation and Related Work (Q2–Q3)

Q2. Scaling Latent and Representation Dimensions

We further increased the latent and representation dimensions jointly, covering (8, 4), (16, 8), and (32, 16).

Using a unified experimental setup, we obtained the following test results:

Latent dimension / representation dimensionMethodTarget-subspace errorLinear-structure error (DlinD_{\mathrm{lin}})
8 / 4Gaussian5.26925.2706
8 / 4Gaussian + ZCM0.18550.1864
16 / 8Gaussian7.39847.4095
16 / 8Gaussian + ZCM0.45770.4603
32 / 16Gaussian8.54728.5885
32 / 16Gaussian + ZCM0.83140.8445

On the challenging competition spectrum with d=32d = 32 and m=16m = 16, the target-subspace distances are 8.5472 for Gaussian and 0.8314 for fixed closure. The corresponding DlinD_{\mathrm{lin}} values are 8.5885 and 0.8445. The improvement therefore remains substantial at higher dimensions.

Q3. Positioning Relative to Prior JEPA and Latent World-Model Work

We will add these three papers, together with recent work on JEPA-based latent world models, and clarify their relationship to our work.

Sobal et al. (2022) showed that JEPA’s preference for temporal predictability can lead a model to rely on static backgrounds. They demonstrated this shortcut and provided a theoretical explanation. PLDM (Sobal et al., 2025) studies how to learn action-conditioned latent dynamics from reward-free offline trajectories. It combines latent prediction and related objectives to obtain representations suitable for downstream planning.

Our work goes beyond asking which slow variables a model prefers. We investigate a deeper representational-structure question. When latent variables have different transition rates and the representation dimension is smaller than the latent dimension, both the latent variables and their nonlinear functions compete for limited representation slots. We characterize this competition through the Hermite spectrum. We further prove that, even when the output strictly satisfies the Gaussian distribution constraint, multiple independent Gaussian representation coordinates may still be generated by the same latent source. This provides a theoretical way to determine whether a finite-dimensional representation genuinely uses distinct sources of latent information.

LeWorldModel (Maes et al., 2026) uses next-embedding prediction and SIGReg Gaussian regularization to achieve stable end-to-end pixel-based world-model learning. However, Gaussian regularization primarily constrains the marginal output distribution. We characterize the source-reuse ambiguity that this constraint cannot rule out and introduce the ZCM constraint to address it.

These three works advance JEPA research from different perspectives, but do not directly answer whether different coordinates in a finite-dimensional representation cover distinct sources of latent information. Our contribution addresses this theoretical gap. Even when the prediction objective and Gaussian regularization have avoided conventional forms of collapse, a finite-dimensional representation may still reuse the same latent source across different coordinates. We characterize this ambiguity and provide a method for eliminating it under explicit conditions and recovering linear latent structure.

After recovering the linear selection space, task-relevance methods could select latent variables that better match the task, potentially reducing interference from slow background factors. Combining these approaches is an interesting direction for future work.

We believe that our theoretical analysis deepens the understanding of representation selection in JEPA and provides a theoretical basis for understanding and designing predictive representations with limited capacity.

Reviewer_PZ8gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how LeJEPA allocates limited representation dimensions among latent variables and their nonlinear functions. It shows that predictable nonlinear functions can displace faster latent variables, and that Gaussian output constraints alone do not prevent multiple coordinates from reusing one latent source. The authors propose a zero conditional-mean (ZCM) transition-closure constraint, which, under the stated Gaussian and observability assumptions, supports linear latent recovery. They also provide stability results and synthetic experiments.

Critical strengths

The paper addresses a meaningful question: what information is selected when representation capacity is smaller than the latent dimension? The distinction between identifying linear structure and selecting a target subspace is clear and useful.

The Hermite-spectrum analysis provides an intuitive account of why nonlinear functions of slowly evolving variables can displace faster linear variables.

The source-reuse construction highlights an important limitation of Gaussian marginal constraints: Gaussian outputs need not represent distinct latent sources.

Critical weaknesses

The pixel experiment is not clearly covered by the main theorem, which assumes that observations preserve all latent information. The paper should clarify whether the pixel rendering is invertible and distinguish observable-space closure from latent-space closure.

The empirical evidence does not adequately verify the quantitative assumptions of the approximate recovery theorem. In particular, the paper should report independent estimates and uncertainty for the closure error, Gaussianity/fourth-moment error, and predictability floor.

The restricted ZCM critic can underestimate the true closure defect. Without stronger-critic or known-transition audits, the reported improvement may reflect critic underfitting rather than genuine closure.

Key experimental details and metric definitions are insufficiently specified for reproducibility. The manuscript also appears to omit ICLR 2027’s required AI-use disclosure.

Questions for authors

1.Is the pixel observation map information-preserving? If not, how do the authors address the gap between observable closure and latent closure described in Appendix D.2?

2.Does Figure 3C satisfy the numerical conditions of Theorem 4, including ε < b²/8? Please report independent estimates and uncertainty for these quantities.

3.How do the authors verify that the ZCM critic is expressive and optimized enough not to underestimate the closure defect? Was the learned representation audited using a stronger critic or the known synthetic transition?

2 comments
AuthorsOpenReview ↗

Observation Identifiability and Quantitative Recovery Conditions (Q1–Q2)

Q1. Observation Identifiability and the Pixel Experiment

The original manuscript did indeed lack a proof that the pixel observations satisfy the information-preservation assumption. We have now added a proof of invertibility for the original soft-disk renderer, together with a control experiment that has an explicit inverse. These additions address the observation assumption and do not change the main theorem.

For the original 64×64 pixel renderer, we prove global injectivity in the ideal real-valued model. The proof uses three fixed 2×2 pixel blocks. If two parameter settings produced the same image, they would force the second derivative of a distance-transform function to vanish on three distinct intervals. Unless the function is the identity, it can have at most two zeros. The parameters must therefore coincide. Combining this result with the invertible tanh transformation between the physical parameters and the four latent variables shows that the image preserves all four latent variables.

We also numerically inverted the samples used in the pixel experiment. Given only the image and the known rendering rules, and without using the ground-truth latent variables for initialization, the inversion accurately recovers all four latent variables.

The new analytic-inverse control uses four fixed orthogonal patterns on a 64×64 image and arranges them as the columns of a matrix B with orthonormal columns. We mix the four latent variables Z into the image as X=sigmoid⁡(32BZ)X = \operatorname{sigmoid}(32 BZ), where BTB=IB^{\mathsf T}B = I. Applying the logit transform to the pixels and projecting onto the four patterns recovers Z exactly as Z=BTlogit⁡(X)/32Z = B^{\mathsf T}\operatorname{logit}(X)/32. Thus, in the idealized setting, the observation preserves all information.

The original pixel experiment reports target-subspace distances of 1.475±0.712 (Gaussian) and 0.150±0.015 (Gaussian+Closure). The new analytic-inverse control gives independent fitting/test target-subspace distances of 1.430±0.977 and 0.053±0.011, respectively. These results support the conclusion that, with information-preserving observations in a controlled setting, adding closure substantially improves target-subspace recovery.

For general real-world scenarios, we do not assume that the observation is strictly invertible. We have added the exact decomposition

εZ=εX+∥E[h(Z′)∣Z]−E[h(Z′)∣X]∥L22.\varepsilon_Z = \varepsilon_X + \left\| \mathbb{E}[h(Z^{\prime}) \mid Z] - \mathbb{E}[h(Z^{\prime}) \mid X] \right\|_{L^2}^{2}.

Thus, preserving all latent-state information is sufficient to eliminate the second term. Predictive sufficiency of the information retained by a particular representation is also sufficient. If neither condition has been verified, a small observable-space closure error does not by itself imply a small latent-space closure error. We will explicitly distinguish the experimental controls with a rigorous invertibility proof from the empirical results on real data, and restrict the applicability of Theorem 3 to its original assumptions. This revision does not change the theorem; rather, it makes the theoretical treatment more complete and clarifies the relationship between the theorem’s assumptions and the experimental validation.

Q2. Quantitative Conditions for Stable Recovery

Figure 3C satisfies the condition ε<b2/8\varepsilon < b^2/8 that you asked about. According to the analytic calculation in Appendix F.8, b=0.7b = 0.7 and ε\varepsilon is at most 0.002025, whereas b2/8=0.06125b^2/8 = 0.06125. The inequality therefore holds throughout Figure 3C. These are analytic values independent of the trained critic, and we will add the calculation.

This example supports the stability mechanism under study. To examine whether the conditions hold for actual trained networks, we independently measured the following three quantities:

MethodClosure error ε\varepsilon ↓Fourth-moment discrepancy Δ4\Delta_4 ↓Predictability floor bb
Gaussian0.0258480.5080.873
ZCM0.0016050.5440.204

The closure residual decreases substantially, while the fourth-moment discrepancy is not noticeably lower than that of the Gaussian baseline. All five seeds satisfy ε<b2/8\varepsilon < b^2/8, although this does not strictly establish that the condition holds uniformly throughout the trained networks.

AuthorsOpenReview ↗

Critic Validation and Reproducibility (Q3–Q4)

Q3. Critic Capacity and Independent Audit

To test whether a small critic objective is a useful optimization proxy during training, we trained new critics with different configurations. This allowed us to examine critics with different expressive capacities. The results are as follows:

Critic width / updates per stepTarget recovery error ↓Independent closure error ε\varepsilon ↓
No ZCM2.582640.02356
32 / 30.940220.01268
128 / 30.060500.00588
512 / 30.060690.00596
128 / 11.787090.01614
128 / 100.050510.00230
128 / 201.629740.02202
128 / 402.272070.02683

We tested capacity sensitivity by varying the width of the critic used during training. Under the same update budget, increasing the width to 128 leads to a clear improvement, while increasing it further to 512 brings no additional benefit.

With the critic width fixed, we compared the number of updates per step. Three and ten updates both yield substantial improvements, whereas performance deteriorates when the number of updates is increased further. We also tried deeper MLPs, residual networks, and cross-evaluation approaches, but none outperformed the setting with ten updates. The independent audit also gives residuals of 0.02356 for Gaussian and 0.00230 for ZCM, consistent with the 128/10 configuration.

Theoretically, for a fixed representation, the population dual objective satisfies

J(c)=εX−∥E[r∣X]−c(X)∥L22.J(c) = \varepsilon_X - \left\| \mathbb{E}[r \mid X] - c(X) \right\|_{L^2}^{2}.

When relating a finite critic to latent-space closure, one must also account for the critic approximation gap, so some structural underestimation is indeed possible. Nevertheless, the experiments provide empirical evidence that, with a reasonable critic configuration, the critic is adequate as a training surrogate and further optimization yields only marginal benefits.

Q4. AI-Use Disclosure and Reproducibility

The original submission includes an AI USE STATEMENT on page 10 describing the scope of AI assistance and the authors’ responsibility.We will make this section more prominent in the revised layout.

For reproducibility, we have further checked the original experimental records and organized the detailed experimental settings and metric definitions. We will provide more detailed definitions in the appendix. We have also provided part of the experimental code and results, and will add the materials associated with the newly added experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.