acceptodds
Under review as a conference paper at ICLR 2027

NextWM: Non-Extensive Joint Embedding Predictive Architectures

Abstract

World models that plan in a learned latent space rely on a distributional regularizer to prevent representation collapse, and the target of that regularizer shapes the geometry in which the planner measures progress toward a goal. Current latent prediction world models fix this target to a Gaussian distribution . We ask what happens when that choice is relaxed. NextWM (Non-extensive World Model) replaces the Gaussian with a heavy-tailed -Gaussian from Tsallis statistics, which recovers the Gaussian as a special case, while keeping the architecture, prediction loss, characteristic-function estimator and planner unchanged. Across four manipulation and navigation environments, NextWM predicts future latents more accurately than a trained world model with the Gaussian target on every task, yet its planning success still need further investigation. Our results shows the potential benefits of non-extensive targets for latent prediction world models, and motivates future work on how the geometry of the latent space affects planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Planning, Self-supervised Learning, Tsallis Statistics, World Models

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_Gb7WOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the possibility of relaxing the isotropic Gaussian regularization in latent world model training. To achieve this, they propose QCFReg, a regularization method that replaces the Gaussian distribution with a heavy-tailed Student-t distribution. Following LeWM, this paper compares models trained with SIGReg and QCFReg in success rate, latent geometry, rollout MSE, and different horizons. However, my concerns are: 1) this paper does not show why a heavy-tailed distribution is needed, 2) each task needs a different set of hyperparameters, which obscures the multi-task possibility, and 3) the success rates do not support the claim that QCFReg is better than SIGReg.

Critical strengths
  1. The comparison and analysis are clear and comprehensive.
  2. The motivation is clear.
Critical weaknesses
  1. The only difference between QCFReg and SIGReg is the target distribution, but there is no clear evidence of why this change is needed. In fact, some results show that this change hurts training simplicity and the possibility of multi-task prediction.
  2. No long-horizon and OOD tests are included. With the nearly saturated performance on these tasks, it is necessary to provide more tests in real-world scenarios.
Questions for authors
  1. I want to see clear evidence that this change to the target distribution is needed, not speculation.
2 comments
AuthorsOpenReview ↗

Response to Reviewer

We sincerely thank the reviewer for their remarks and comments. Your comments were very insightful and are shaping the new version of the paper.

Following, we tried to address your questions based on the new paper version. To help the review process, main changes are highlighted in blue in the new uploaded pdf.

What evidence shows that changing the regularizer target is necessary?

None of the previous evidence establishes necessity. We removed that claim and retain the Gaussian target in the main method. q-gaussian is a matched-variance control. Table 1 shows prediction improvements with qq-loss under either target at 10% swapping.

Different hyperparameters for each task

The Push-T comparison fixes q=1.25q=1.25, batch-median scaling, mean-weight normalization, target variance four, regularizer weight 0.09, and training budget, without per-arm tuning. We have a future work to address a learnable qq during the optimization, avoiding a fixed qq value.

Longer horizons, out of distribution, or in real-world scenario

This is interesting point, that motivated us to add a new section in the appendix. The paper reports prediction through horizon five, but broader horizon generalization, environmental distribution shift, and real-world robustness remain untested. Short fault pilots (paper Figure 6) improve on swaps and black frames but worsen on clean, frozen-frame, and mismatched-action data. We are planning real robot manipulation tasks for a future paper.

Table R3.3. Short fault pilots: one-step latent NMSE. Push-T, 6,000 steps, one seed per arm; Student-tt target, q=1.25q=1.25, mean-weight normalization. Lower is better.

Training conditionFault fractionMSE NMSE₁qq-loss NMSE₁
Clean0%0.0390.044
Swapped targets30%0.1650.061
Black frames30%0.1550.113
Frozen frames30%0.0510.079
Mismatched actions30%0.0410.062

I appreciate the pollution experiment, as it provides evidence supporting the effectiveness of q-loss. However, given the limited room for improvement in success rates on in-domain and short-horizon tasks, it remains unclear whether the proposed method makes a sufficiently significant contribution to the community. Therefore, I will maintain my current score.

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

This paper proposes replacing the Gaussian target in LeWM's distributional regularizer with a heavy-tailed q-Gaussian from Tsallis statistics. Experiments show slightly higher planning success on Push-T but lower success rates than LeWM on the other three environments, with no demonstrated advantage in collapse prevention. Although the proposed method achieves lower latent prediction errors across all four environments, these errors are measured in different learned latent spaces and are therefore not directly comparable as evidence of improved dynamics modeling.

Critical strengths

None.

Critical weaknesses
  1. Limited novelty and insufficient motivation. The main methodological contribution is replacing the Gaussian target in LeWM's existing distributional regularizer with a heavy-tailed q-Gaussian, while leaving the rest of the framework unchanged. The paper provides little theoretical or empirical justification for why heavy-tailed latent representations should be preferable for robotics or control tasks. The methodological novelty is therefore limited.

  2. Lack of demonstrated improvements in planning or collapse prevention. The experiments do not establish that the proposed q-Gaussian regularizer is more effective than the original Gaussian regularizer. Planning success improves slightly on only one of the four environments and decreases on the other three, with no statistically established improvement. The representation diagnostics also do not demonstrate superior collapse prevention or identify a geometry that benefits planning.

  3. Limited comparability of latent prediction errors. The prediction errors in Section 4.5 are measured in different learned latent spaces and are therefore not directly comparable as evidence of superior dynamics modeling.

  4. Missing essential experimental controls. The comparison changes both the tail shape and variance of the target distribution (from 1 for the Gaussian target to 4 for the q-Gaussian target), making it impossible to isolate the effect of heavy tails. Moreover, the Gaussian baseline relies on released checkpoints rather than matched retraining, and only one training seed is used for each new model, further limiting the attribution and reproducibility of the reported results.

Questions for authors

None.

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We sincerely thank the reviewer for their remarks. Your points where in the right direction pointing that the q-Gaussian target adds near zero value to the planning or error. Based on this, we reframe our experiments to explore how a q-Gaussian prediction error helps in a collapsed/collapsed training data. And our current results start to indicate an advantage of our method.

Following, we tried to address your questions based on the new paper version. To help the review process, main changes are highlighted in blue in the new uploaded pdf.

Motivation of the proposed change

Based on the revised approach, our hypothesis concerns sensitivity to corrupted training transitions. We apply the a q-Gaussian prediction loss to JEPA-style training and analyze its influence under fixed-cutoff assumptions. Our results in Table R2.1 motivates this direction; it does not establish general drift robustness or validate the batch-adaptive implementation theoretically

Table R2.1. Sensitivity to target corruption with the Gaussian target fixed. Derived from revised paper Table 1; lower NMSE is better.

Prediction lossnn clean / corruptedClean NMSE₁10% swaps NMSE₁Corrupted / clean
MSE1 / 20.00150.01258.51×
qq-loss1 / 10.00130.00171.34×

Do the results establish better planning or collapse prevention?

Planning success is higher under 10% swapping: 92.7% to 96.0% with the Gaussian target (+3.3 percentage points) and 87.8% to 91.6% with Student-tt (+3.8 points. Table below and Table 1 in the udpated paper). These observed gains support the hypothesis, but limited training replication is running to confirm the statistical gain across multiple seeds. Representation-collapse prevention is a separate claim that planning success alone does not establish, but we can address in a extra experiment or future research.

Planning success from revised paper Table 1.

Target / trainingnn MSE / qqMSE success (%)qq-loss success (%)Difference (pp)
Gaussian / clean1 / 194.094.7+0.7
Gaussian / 10% swaps2 / 192.7 ± 2.896.0+3.3
Student-tt / clean2 / 289.3 ± 0.089.7 ± 2.4+0.3
Student-tt / 10% swaps3 / 387.8 ± 0.491.6 ± 2.1+3.8

Are prediction errors comparable when each model learns a different latent space?

We can say that latent NMSE alone is insufficient because representations differ. We therefore report linear-probe error in common standardized task coordinates, using matched capacity and fitting procedures (Table 1 in the updated paper).

Missing matched baselines and independent training replications

Partly for now. Table 1 reports matched MSE/qq controls and all available seed counts. We are running extra experiments with other robust losses, like Huber, to compare the results. But our current results can indicate the advantage of our in this scenario.

Reviewer_ZLUjOpenReview ↗

Rebuttal Response

Thank you to the authors for the response and revisions. The authors acknowledge the limitations of the original q-Gaussian regularization approach and have substantially changed the paper's central focus from heavy-tailed regularization to a robust prediction loss under corrupted training data. However, the new contribution remains limited, with full-budget experiments restricted to Push-T with synthetic target corruption and insufficient comparisons with existing robust losses. Therefore, I maintain my original rating of Clear Reject.

Reviewer_KPf2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NextWM replaces LeWM’s Gaussian characteristic-function target with a q-Gaussian target while retaining the model design, prediction objective, and planner. The main setting, q=1.5, β=0.5q=1.5,\ \beta=0.5, corresponds to a scaled Student-t distribution with variance four. The paper evaluates planning, autoregressive latent prediction, and representation geometry across four environments.

The evaluated NextWM checkpoints have lower variance-normalized prediction error at every tested horizon. At the standard planning budget, their mean success exceeds released LeWM on Push-T but is lower on the other three tasks.

Critical strengths
  1. A technically coherent extension with a clear implementation path.

Sections 3.2-3.5 specify the target, its scale, the numerical estimator, and the actual training gradient. The Gaussian limit is explicit. Appendix F correctly distinguishes the one-dimensional q-Gaussian from the multivariate Student-t construction, and correctly notes that the main target has finite variance but no finite fourth moment. These details support a usable alternative regularizer without introducing a different planning objective.

  1. Useful evidence that prediction scores and planning scores need separate assessment.

The rollout probe uses common observation windows, recorded actions, autoregressive prediction, variance normalization, and a persistence reference. Thus, the comparison is more informative than reporting raw latent MSE alone. The consistent prediction-error ordering across four tasks contrasts with the mixed planning ordering.

Critical weaknesses
  1. The experiments do not establish what the proposed target change causes.

The main comparison uses released LeWM checkpoints with unmatched training, while each new NextWM model has only one training seed. Moreover, the principal targets change variance as well as shape: the Gaussian reference has target variance one, whereas q=1.5,β=0.5q=1.5,\beta=0.5 has variance four. The q=1.25q=1.25 reference does not resolve this because it uses yet another shape-scale combination. The paper itself correctly explains that scale changes can alter both prediction-loss magnitude and characteristic-function matching during training.

Consequently, neither the prediction differences nor the planning differences can be attributed specifically to the heavy-tailed target. Evaluation seeds and bootstrap intervals cannot resolve training-level uncertainty. Section 5 acknowledges these issues, but acknowledging them does not answer the central experimental question. A matched Gaussian control at the same target variance is more important here than adding further environments or unrelated baselines.

  1. The main prediction result does not yet establish better task-relevant dynamics prediction.

Variance normalization removes uniform scale, so I do not regard Table 2 as a simple raw-MSE comparison error. However, each model predicts its own learned representation. A representation can become easier to predict by changing which information it retains; persistence-relative skill does not establish that the retained information is equally useful for control. Table 4’s lower effective rank makes this worth checking, but does not itself demonstrate lost information or collapse.

Further, the validation windows are not guaranteed to be trajectory-disjoint from training for both models. Candidate-action ranking fidelity is also unmeasured. The evidence therefore supports lower error in each checkpoint’s own latent metric, but not a generalization advantage or an explanation of how target-induced geometry changes planning.

Questions for authors

a) Can a Gaussian target with variance four, q=1,β=1/8q=1,\beta=1/8, be compared with the main NextWM setting under matched training, initially on Push-T and with independent training seeds? This is a focused control for the claimed benefit, not a request for the full proposed 72-run study. Persistence of the effect under this control would materially strengthen the contribution.

b) Table 3 reports 95.3%95.3\% mean success for retrained LeWM versus 93.3%93.3\% for NextWM, but notes that the retrained reference used unseeded CEM. Can that existing checkpoint be evaluated with the common seeded planner and rollout probe, with its training settings specified? The current figures do not establish that it is better, but this comparison could substantially change the interpretation of the Push-T result.

c) Can you identify a trajectory-disjoint evaluation set for both checkpoints and test whether the advantage persists there? On that set, one shared task-space measure, such as future object-position error from matched-capacity probes, would help distinguish better dynamics prediction from easier-to-predict representations. This would address the scope limits already stated in Section 5.

1 comment
AuthorsOpenReview ↗

Response to Reviewer

We sincerely thank the reviewer for their remarks.

Following, we tried to address your questions based on the new paper version. To help the review process, main changes are highlighted in blue in the new uploaded pdf.

Question 1

We really appreciate this question. Table 1 of the revised paper compares Gaussian and Student-tt targets at variance four with matched architecture, schedule, and episode split. Gaussian replication remains incomplete. The results support studying the prediction loss, but establish neither a heavy-tail advantage nor a complete shape-versus-scale comparison.

Table R1.1. Matched-variance targets under 10% swapping. From revised paper Table 1: variance four, 155,000 steps; nn counts training seeds.

TargetLossnnNMSE ↓\downarrowState RMSE ↓\downarrowSlope
GaussianMSE20.01250.3320.896
Gaussianqq-loss10.00170.2380.998
Student-ttMSE30.01260.4200.896
Student-ttqq-loss30.00350.3010.998

Question 2

The reported comparison does not include a seeded re-evaluation of that historical LeWM checkpoint. We removed the associated planning-advantage claim. Table 1 instead compares newly trained MSE and qq-loss controls using the same seeded planner.

Question 3

Prediction uses 256 held-out episode windows and matched linear probes fitted on training episodes. Historical planning starts and state/action scaling used the full dataset. Strict evaluation isolation requires training-only preprocessing and held-out planning starts.

Table R1.3. Shared task-space prediction error at horizon five. From revised paper Table 1; lower is better.

Target / trainingnn MSE / qqMSE RMSEqq-loss RMSEqq vs. MSE
Gaussian / clean1 / 10.2350.237+0.7%
Gaussian / 10% swaps2 / 10.332 ± 0.0040.238-28.5%
Student-tt / clean2 / 20.300 ± 0.0100.297 ± 0.009-0.9%
Student-tt / 10% swaps3 / 30.420 ± 0.0170.301 ± 0.005-28.4%

Question 4

Thanks for this question. The direct answer is that we are observing mixed resutls from our results, a more detailed evaluation will be necessary or a specific future paper exploring this trade-off in specifc. Task-space probes and planning success provide complementary evidence (paper revised Table 1), but candidate-action ranking fidelity remains unmeasured. Lower latent error alone does not establish better task-relevant dynamics or trajectories in our results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.