acceptodds
Under review as a conference paper at ICLR 2027

How Latent Transition Geometry Matters for Planning in World Models

Abstract

World models support planning by predicting how the environment will evolve under candidate actions. In latent world models, the model is typically trained for one-step prediction, while planning typically requires repeatedly composing the learned transitions to compare candidate futures over a multi-step horizon. However, one-step latent predictive accuracy does not automatically translate to planning success. This raises a question: *what geometric properties of the learned one-step transition are useful for multi-step planning?* We study this question through an action-conditioned *transition-field* view of latent dynamics, using curvature regularization as a controlled geometric intervention. Across four continuous-control tasks, appropriately tuned regularization improves planning, while excessive regularization can sometimes be detrimental. To understand this observation, we decompose the transition field into an affine component and a residual non-affine component. We find that the curvature intervention reduces the field's sensitivity to action primarily by contracting its non-affine component. These sensitivity reductions propagate through the rollout, blurring the distinction among candidate futures and, in turn, altering the planner's decisions. Together, these results suggest that useful latent dynamics should retain sufficient action-induced predictive differences through multi-step propagation to keep candidate futures distinguishable to the planner.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
World Models, JEPA, Transition Geometry, Planning, Curvature Regularization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Bv9tOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how the geometry of learned latent transitions affects multi-step planning in world models. It argues that accurate one-step latent prediction does not necessarily translate into effective planning, because action-induced differences must remain sufficiently distinguishable as transitions are repeatedly composed over a planning horizon.

Using LeWorldModel (LeWM) as the experimental framework, the paper makes three main contributions.

  1. it formulates latent dynamics as an action-conditioned transition field and develops a framework for analyzing how local action responses propagate through multi-step rollouts

  2. through curvature regularization, it shows that trajectory straightening primarily contracts the non-affine component of the transition field, reducing sensitivity to actions and changing the candidate action sequences selected by the cross-entropy method

  3. it identifies three complementary properties of rollout geometry control scale, conditioning, and distance fidelity which help s characterize the relationship between latent dynamics and planning performance

Critical strengths
  1. A useful conceptual distinction between one-step prediction accuracy and multi-step planning effectiveness. The paper highlights that accurate latent transition prediction does not necessarily yield good planning, because the geometry of action-conditioned transitions can influence how action differences propagate through rollouts.This provides a meaningful perspective on world-model evaluation beyond one-step prediction metrics and offers a potentially useful framework for understanding planning failures.

  2. Mechanistic analysis connecting latent transition geometry to downstream planning behavior. Rather than reporting only aggregate performance, the authors investigate how curvature regularization affects control scale, conditioning, and distance fidelity, and trace its influence viaclosed-loop performance. Experiments across four continuous-control tasks are shown.

Critical weaknesses

1.The evaluation is restricted to LeWorldModel (LeWM) and four continuous-control tasks: TwoRoom, PushT, Reacher, and Cube. This limits the strength of the paper's broader claims about world-model-based planning.

  1. Limited evidence that the proposed geometric properties provide a generalizable causal explanation of planning performance. The analysis links curvature regularization to changes in action sensitivity, CEM elite selection, and closed-loop performance, but the observed non monotonic effects also indicate that the relationship between transition geometry and planning success is complex and task dependent.
Questions for authors

1.How consistently do these properties explain planning performance across the four tasks, and do the authors have quantitative evidence that they predict performance beyond the specific curvature-regularization experiments presented?

2.It would strengthen the paper to evaluate whether the proposed analysis extends to other world-model architectures and planning algorithms. Do the authors have preliminary evidence or a principled argument suggesting that the findings are not specific to LeWM and CEM?

1 comment

Thank you for your constructive feedback and for recognizing the value of our analysis. We have revised the manuscript to clarify the scope of our findings and added evidence across planning solvers and world models.

W1 / Q2. Generalization across planners and world models. Our formulation is based on the differentiable action-to-future map and its relation to the planning objective. In principle, the same analysis therefore applies to other differentiable latent world models and across planning solvers. To examine this, we added comparisons across five planners (CEM, iCEM, MPPI, GD, and Adam) on LeWM, and evaluated DINO-WM on PushT in Appendix F, page 19. Figure 9 shows broadly consistent planning trends across solvers, while Table 8 reproduces the rollout-geometry trend in DINO-WM. These results provide additional evidence that the analysis extends beyond LeWM and CEM.

W2 / Q1. Refined analysis. We have refocused the rollout analysis in Section 6, while clarifying its connection to the original diagnostics. The original control scale measured the Frobenius norm of the covariance-weighted rollout Jacobian, summarizing the overall magnitude of action-induced terminal variation. We now use its squared norm as total response energy and report this energy relative to the base model through RtotalR_{\mathrm{total}}. This makes the degree of overall contraction directly interpretable. The original conditioning measure summarized spectral imbalance across controllable directions, without distinguishing which directions matter for the current goal. We therefore focus the directional analysis on the response along the direction from the predicted terminal state toward the goal, versus the response in orthogonal directions. The diagnostic qq measures how the balance between these two components changes relative to the base model.

The sweeps in Figure 5 show that stronger regularization generally contracts the overall rollout response while relatively preserving its goal-direction component. In beneficial regimes, this pattern accompanies improved planning. Under excessive regularization in PushT, however, the sharp decrease in RtotalR_{\mathrm{total}} outweighs the increased relative retention measured by qq: goal-direction responses also shrink, potentially weakening the distinctions needed to compare candidate actions. This helps explain why planning deteriorates even as qq continues to rise. Figure 4 schematically illustrates these two regimes. We interpret these quantities as local geometric diagnostics, rather than necessary or sufficient conditions for planning success. The useful response scale depends on the task and planner, and establishing generalizable predictive criteria across these settings remains a direction for future work.

Reviewer_Ha9dOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how the geometry of learned latent transitions influences multi-step planning in world models. The authors formulate action-conditioned latent dynamics as a transition field and analyze how local action sensitivity propagates through repeated model rollouts. Using curvature regularization as a controlled intervention, the paper examines changes in transition geometry and their effects on planning across four continuous-control tasks. The analysis further decomposes the transition field into affine and non-affine components, showing that curvature regularization primarily contracts the non-affine action response. Through targeted interventions and CEM candidate-selection experiments, the authors investigate how these changes affect planning decisions and closed-loop performance. Finally, the paper studies rollout-level conditioning, control scale, and distance fidelity to characterize the relationship between latent transition geometry and planning effectiveness.

Critical strengths
  1. The paper provides a systematic investigation of how curvature regularization affects planning across four continuous-control tasks. The experiments cover multiple regularization strengths and three independent training seeds, revealing that the same geometric intervention can improve planning in some settings while substantially degrading performance in others. This task-dependent behavior provides useful evidence that reducing latent trajectory curvature is not universally beneficial and motivates a more careful examination of how learned transition geometry influences downstream planning.

  2. Beyond reporting correlations between geometric measurements and planning performance, the paper explicitly modifies the non-affine action-dependent component of the transition field and examines its effects on CEM candidate selection and closed-loop control. The results show that changing this component alters the selected elite action sequences and can significantly affect planning success. These interventions strengthen the empirical analysis by demonstrating that the identified non-affine response contains information actively used by the planner, making the mechanistic investigation more informative than a purely observational study.

Critical weaknesses
  1. The paper builds heavily on existing ideas in latent trajectory geometry and action-sensitive world modeling. In particular, Temporal Straightening (Wang et al., 2026) has already investigated curvature regularization and its relationship with latent distance fidelity, planning conditioning, and downstream control performance, while Delta-JEPA (Zhang et al., 2026) emphasizes preserving action-dependent differences in learned latent transitions. The main additional contribution of this paper is therefore the affine/non-affine decomposition of action sensitivity and its connection to candidate selection during planning. Although this analysis is interesting, the current evidence does not sufficiently establish how much new understanding it provides beyond the previously identified importance of trajectory geometry and action sensitivity. Stronger evidence isolating the explanatory contribution of the proposed decomposition would help justify its significance.

  2. The paper demonstrates that scaling the non-affine action response changes CEM elite selection and closed-loop performance, but this intervention also modifies the predicted transitions for candidate actions away from the proposal center. Consequently, the observed performance changes may partly result from altered prediction accuracy rather than the proposed geometric mechanism alone. Preserving the prediction at the proposal center does not eliminate this confounding factor. Furthermore, although the regularization sweeps use three independent training seeds, the central decomposition, action-sensitivity measurements, and intervention analyses are performed using only one selected checkpoint for each task and regularization strength. Without examining whether these findings remain consistent across training seeds and controlling for changes in prediction error, the claimed mechanism is not yet sufficiently established.

Questions for authors
  1. How consistent are the proposed geometric mechanisms across independently trained checkpoints? The main regularization experiments use three training seeds, but the key affine/non-affine decomposition and intervention analyses rely on one selected checkpoint per task and regularization strength. Could the authors repeat the principal action-response attribution and intervention experiments across the three existing training seeds, reporting variability in non-affine contraction, CEM elite selection, and planning performance? Establishing that these effects are consistent across checkpoints would substantially strengthen the reliability of the proposed explanation.

  2. Can the authors distinguish the effects of action-response geometry from changes in prediction accuracy? The intervention in Eq. (19) modifies the non-affine response while preserving the prediction at the proposal center, but it may also introduce prediction errors for other candidate actions. Could the authors report one-step and multi-step prediction errors across the existing response-scaling experiments, alongside CEM selection and planning success? Additionally, could they include a control perturbation matched in prediction error but not explicitly targeting the non-affine action response? Such evidence would help determine whether the observed planning changes are specifically caused by the proposed geometric mechanism rather than a more general deterioration in model accuracy.

5 comments

Thank you for the helpful feedback. We have revised the manuscript to clarify the contribution and strengthen the evidence for the proposed mechanism.

W1. Clarified contribution. Our contribution goes beyond observing that curvature regularization affects planning: we identify where the action response changes and directly test how that component affects the planner. Section 4 attributes most of the curvature-induced reduction in action-response energy to the non-affine component. Section 5 directly scales its action-induced differences in frozen models and measures changes in CEM selection and closed-loop performance. Section 6 connects these findings to overall rollout contraction and relative goal-direction retention, while Section 7 provides complementary encoder-attention observations. Together, these analyses suggest that planning depends on both the scale and directional distribution of action-induced differences: moderate contraction can be beneficial, whereas excessive contraction can weaken candidate distinctions despite greater relative goal-direction retention.

W2 / Q1. Cross-seed consistency. We extended the principal LeWM action-sensitivity, non-affine attribution, and response-scaling experiments to three seeds, with the evaluation and aggregation protocol specified in Appendix A, page 14, line 731. The updated Figure 2, Table 2, and Figure 3 report variability across seeds and support the same qualitative conclusions: contraction is concentrated in the non-affine response, and directly scaling this response changes CEM selection and planning. As summarized on page 8, line 421, Table 7 further supports the rollout-geometry findings across three seeds at the beneficial and over-regularized settings.

W2 / Q2. Prediction-error analysis and controls. We added one-step and five-step prediction-error analyses in Appendix D, page 17, line 883. At β=0\beta=0, increased errors accompany worse planning, whereas under partial attenuation, error changes do not consistently track planning outcomes. On the unregularized TwoRoom checkpoint, we further constructed two oracle controls that scale prediction residuals toward encoded environment observations, calibrated to match either the one-step or five-step mean error at β=0.75\beta=0.75. Both match the mean errors at both horizons within 2.5%2.5\%, yet elite overlap remains approximately 0.860.86, versus 0.6610.661 under non-affine response scaling. Thus, matching mean prediction errors alone did not reproduce the same change in CEM candidate selection in this control.

Thank you for the additional experiments. The three-seed analyses address my concern about consistency across checkpoints. However, the reported error-matched controls are limited to one TwoRoom checkpoint and demonstrate differences in elite selection, leaving their connection to closed-loop performance unclear. Matching mean prediction errors also does not control their direction or variation across candidate actions, so my concern about separating geometry from prediction accuracy remains only partially addressed. Consequently, I am still not sufficiently convinced of the additional explanatory value of the affine/non-affine decomposition. I therefore maintain my score.

Thank you for asking how the error-matching result carries over to closed-loop planning. The original oracle requires future observations for each candidate action sequence, so it is limited to offline evaluation. For the closed-loop experiment, we draw on the idea of bias compensation in MPC (Pannocchia and Rawlings, Disturbance Models for Offset-Free Model Predictive Control, 2003). On a fixed batch of offline trajectories, we fit two constant additive offsets to the β=0.75\beta=0.75 predictions, one targeting the β=1\beta=1 one-step MSE and the other targeting its five-step MSE. We fix each offset before evaluation and report both MSEs on held-out trajectories. All four tasks use frozen checkpoints trained with λ=0\lambda=0.

TaskSettingOne-step MSEFive-step MSEPlanning success (%)
TwoRoomβ=1\beta=152.8140.084
β=0.75\beta=0.7544.399.990
β=0.75\beta=0.75 + one-step match51.4100.892
β=0.75\beta=0.75 + five-step match44.3139.096
PushTβ=1\beta=12.311.286
β=0.75\beta=0.753.013.888
β=0.75\beta=0.75 + one-step match3.113.5--
β=0.75\beta=0.75 + five-step match3.014.0--
Reacherβ=1\beta=122.074.882
β=0.75\beta=0.7523.171.786
β=0.75\beta=0.75 + one-step match23.372.180
β=0.75\beta=0.75 + five-step match23.174.188
Cubeβ=1\beta=119.380.470
β=0.75\beta=0.7522.686.170
β=0.75\beta=0.75 + one-step match22.785.7--
β=0.75\beta=0.75 + five-step match22.686.372

The clearest comparison is between β=1\beta=1 and β=0.75\beta=0.75 with five-step calibration. On TwoRoom and Reacher, the calibrated five-step MSE is within about 1% of the β=1\beta=1 value, while closed-loop success increases from 84% to 96% and from 82% to 88%, respectively. In Reacher, the calibrated model also has slightly higher one-step MSE than the original model. Scaling the non-affine action response alone, while holding the affine action term fixed, already improves success to 90% and 86% and lowers five-step MSE on these two tasks.

For PushT, we also tried β=0.5\beta=0.5 and β=1.25\beta=1.25 without finding a match. A dash indicates that we skipped closed-loop evaluation because the error at the targeted horizon remained more than 10% from the β=1\beta=1 value.

We agree that error direction and variation across candidates can still differ after calibration. The observed differences in closed-loop performance show that mean five-step MSE alone does not explain the planning gains across these conditions. More broadly, these results suggest that combining inference-time response scaling with offline error calibration could improve planning without retraining the world model.

We appreciate this suggestion. It led us to test the effect in closed loop and to find a result that the offline oracle could not reveal. We will incorporate these results and the calibration details into the next manuscript revision, along with the other updates from the discussion.

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

The paper is an analysis paper asking which geometric properties of a learned one-step latent transition makes it useful for multi-step planning. It conducts experiments only on top of LeWM, with temporal straightening as the geometry intervention method. Across four environments, its findings are: 1. tuning the weight of straightening helps planning, but how much is task-dependent. 2. Geometric regularization (curvature) can decrease action sensitivity of the world model. 3. They proposed three quantities that are used to explain the different regimes, including conditioning, control scale, and distance fidelity.

Critical strengths
  • The problem of understanding how latent transition geometry matters for planning is relevant.
Critical weaknesses
  • The scope of this investigation is only constrained to LeWM, and only constrained to one way (temporal straightening) to intervene on transition geometry. The paper provides no signs whether the takeaways will generalize to other classes of world models or transition geometry intervention methods.
  • The paper is poorly written. The whole paper reads like an experiment log instead of a paper. It contains large chunks of unnecessary definitions in order to show a simple takeaway.
  • The paper only concerns planning with CEM, which is a sampling-based method that is not directly affected by the transition geometry of the world model. The paper applies temporal straightening, which shows the most improvement for Gradient Descent planning in the original paper. The paper contains no discussion or experiments on GD planning at all.
Questions for authors
  • Could you discuss the generality of the takeaways of the paper, is it confined to only LeWM with temporal straightening?
  • Could you include the same analysis on other classes of world models (e.g. the original temporal straightening is done on DINO-WM) and other geometry shaping methods?
1 comment

Thank you for the feedback. We have revised the presentation and broadened the empirical evaluation.

W2. Clarified presentation and contribution. We reorganized the manuscript around how transition geometry changes the action-induced differences used to compare candidate futures, removing several auxiliary metrics and streamlining the definitions and derivations. Section 4 identifies the non-affine component as the main source of curvature-induced contraction. Section 5 directly scales its action-induced differences in frozen models and measures changes in candidate selection and planning. Section 6 connects these findings to overall rollout contraction and relative goal-direction retention. Section 7 connects this analysis to the learned representations through qualitative observations of how encoder attention shifts across regularization regimes.

Together, these analyses clarify why accurate one-step prediction alone does not ensure effective multi-step planning: planning also depends on how action-induced differences propagate through rollouts and affect the predicted goal distances used to compare candidates. This provides a perspective on world-model evaluation beyond prediction error and helps explain why moderate contraction can improve planning while excessive contraction can weaken useful candidate distinctions.

W1 / W3 / Q1 / Q2. Additional experiments and scope of generalization. We expanded the evaluation to five planners (CEM, iCEM, MPPI, GD, and Adam) on LeWM, and extended the rollout-geometry analysis to DINO-WM on PushT, including planning with both CEM and GD. These results are reported in Appendix F, page 19. Figure 9 shows broadly consistent planning trends across sampling- and gradient-based solvers, while Table 8 reproduces the overall contraction and relative goal-direction retention trend in DINO-WM.

Although CEM does not use model gradients, it ranks candidate action sequences by their predicted costs. Changes in the action-to-future map can therefore affect its selection. Our frozen-model interventions in Section 5, Figure 3, pages 6–7 directly demonstrate changes in CEM elite selection and closed-loop performance across three training seeds.

Our formulation is based on the differentiable action-to-future map and its relation to the planning objective, so the same analysis is applicable in principle across differentiable latent world models and planning solvers. Curvature regularization serves as a controlled intervention for studying this relationship. The frozen-model response-scaling experiments provide a complementary intervention that directly targets the identified non-affine component. We frame the findings as evidence from these interventions, without claiming that all geometry-shaping methods necessarily produce the same effects.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their feedback. We have uploaded a revised manuscript, with key additions highlighted in blue:

Presentation: Reorganized the main analysis and appendices, removed auxiliary metrics, and clarified the connection between action sensitivity, rollout geometry, and planning.
Cross-seed validation: Extended the principal action-response and intervention analyses to three training seeds (Figures 2–3, Table 2), with additional rollout-geometry validation in Table 7.
Prediction-error controls: Added one-step and five-step error analyses and oracle error-matching controls (Appendix D).
Broader evaluation: Expanded LeWM evaluation to five planners and added DINO-WM results on PushT (Appendix F).

Detailed responses are provided under each review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.