acceptodds
Under review as a conference paper at ICLR 2027

Beyond One-Step Accuracy: State-Affine Latent Transition for Reliable Visual Planning

Abstract

Joint-embedding world models enable visual planning by learning action-conditioned dynamics in latent space. Yet they are commonly trained for one-step prediction on encoded states, while planning recursively applies the learned transition to its own predictions. One-step accuracy therefore does not capture how prediction errors propagate under recursive rollout. We decompose multi-step rollout error into the errors introduced at individual steps and their propagation through subsequent transitions. We show that state-affine dynamics are precisely the differentiable transitions with state-independent Jacobians, eliminating the nonlinear propagation residual and making the error propagation operators depend only on the action sequence. Guided by this result, we introduce SALT (**S**tate-**A**ffine **L**atent **T**ransition), an action-conditioned state-affine dynamics model in which the action modulates both the state transformation and the additive update. We train SALT through recursive multi-step rollout supervision, feeding each predicted latent state back into the transition so that training matches how the model is used during planning. Across four visual planning environments, SALT exhibits – higher one-step prediction error than the matched LeWM baseline, yet improves closed-loop success in every environment by percentage points on average. On OGBench-Cube, the fraction of episodes that fail with a sharp rise in model-predicted cost after execution decreases from to .

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
World models, Planning, Error propagation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_fSxNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that one-step prediction error is a poor proxy for how well a latent world model plans because errors compound over a rollout through the state Jacobians. If the transition is affine in the latent state these Jacobians no longer depend on the state. The authors propose SALT on that basis - a small transition model A(c)z+Bc+bA(c)z + Bc + b where the matrix AA depends on the action trained with a 5-step rollout loss. Compared with a reproduced LeWM, SALT has a higher one step error but plans better on four environments (92.7% vs 82.7% average success) with the largest gain on OGBench-Cube (94% vs 67%). It is also much smaller and plans about 3 times faster.

Critical strengths
  1. A simple model with a strong result. A transition that is linear in the state, with a predictor about 16 times smaller plans better and about 3 times faster than a Transformer predictor. On Cube the gain is large and it holds at longer horizons, on Cube-Double, and against the published LeWM number.
  2. The limits are stated plainly. The authors say where their analysis does not apply and the 2×2 ablation (transition structure x training loss) shows what each part contributes.
Critical weaknesses
  1. It is not shown that the affine structure is what helps. SALT trained with the one-step loss is barely better than LeWM so most of the gain comes with rollout training. SALT also differs in size and in its input (one frame instead of three) with no small nonlinear predictor trained the same way to compare against. Since the title credits the affine structure, this is the main issue for my score.
  2. The size of the gain is uncertain. Each method in Tables 1 and 3 is a single trained model, and on PushT and Cube the reproduced LeWM is weaker than the numbers reported in the LeWM paper. Those numbers come from a different protocol, so they are not a direct comparison, but against them SALT loses on PushT and gains less on Cube.
Questions for authors

Clarifications

  1. Can you put the published LeWM numbers in Table 1, explain why your reproduction is lower and report both methods over several training seeds? If the gains hold I would raise my score.
  2. Can you add a small nonlinear predictor (e.g. an MLP with < 1M parameters) that takes the same input as SALT and is trained with the same rollout loss? If it matches SALT the gain comes from size and training rather than the affine structure and the framing should change.
  3. How exactly is LeWM trained with the rollout loss given its three-frame history? Was it tuned for this setting?

Suggestions (not decisive)

  1. Related work: the idea that one-step accuracy does not reflect control quality goes back at least to "Objective Mismatch" (Lambert et al., 2020). Locally linear and Koopman-style latent dynamics for control (PCC, RCE, Koopman MPC) belong in that discussion along with prior rollout-loss results for JEPA predictors (Terver et al., 2026, which the paper cites only for design choices and Lipschitz error bounds; V-JEPA 2-AC, Assran et al., 2025).
  2. Prop. 2 is a known fact. "We show" in the abstract overstates the theory. It works better as motivation for the design.
Reviewer_wxSTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Gradient Penalized Latent Dynamics (GPLD), which regularizes the input Jacobian of DreamerV3’s latent posterior probabilities. The method is motivated by a finite-MDP analysis connecting information sharing between neighboring transition distributions to Jacobian regularization. Experiments show improved learning performance primarily in proprioceptive continuous-control tasks.

Critical strengths
  • The paper provides an interpretable transition-estimation perspective on smoothness regularization.
  • The empirical evaluation includes regularizers, ablations, computational-cost measurements, and diagnostics addressing.
  • Several improvements are supported by substantial seed counts. In particular, GPLD outperforms DreamerV3 on 18 of 22 seeds for walker-run and 16 of 19 for cheetah-run. The nAUC results and earlier attainment of high returns on quadruped provide meaningful evidence of improved interaction efficiency.
  • The paper reports important limitations including weak aggregate gains from pixel observations and additional training cost.
Critical weaknesses
  • The strength of the baseline remains a concern.
  • The incremental contribution and novelty of this research needs clearer separation from existing regularization methods.
  • The paper needs more evidence/analysis for the performance of the proposed mechanism.
Questions for authors
  1. What explains the gap between the rerun and released DreamerV3 results?
  2. Does the GPLD’s gains persist when the stronger baseline performance is reproduced?
  3. Can the walker-run policy improvement be explained by better prediction during early training or on policy-relevant states and actions?
Reviewer_PQcFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the learned latent dynamics that remain effective for planning as prediction error as feedback to the model by decomposing multi-step rollout error into errors introduced in individual steps and their subsequent propagation. The paper also analyzes that the state Jacobian is independent of the latent state if and only if the transition is affine in the state among differentiable transitions. Along with that analysis, they proposed SALT, an action-conditioned transition model that remains affine in the latent state, trained with a recursive multi-step rollout loss. SALT showed higher one step prediction error than the reproduced LeWM but improved closed-loop success in four visual environments, also the authors analyze the failure on the OGBench-Cube.

Critical strengths
  1. SALT seems to be very fast and has fewer parameters (0.70M vs 11.58M predictor parameters, 3.1x faster planning) so that it is cheap to evaluate many candidate action sequences during planning.

  2. The paired failure results on OGBench-Cube show the effectiveness of SALT where SALT succeeds on 43 episodes that LeWM fails and LeWM also fails where SALT fails.

Critical weaknesses
  1. The claim that one step accuracy does not capture how prediction errors propagate comes directly from Prop. 1 because the one step error only measures the new error at each step, while the propagation is the product of Jacobians that the one step loss never trains directly. Prop. 1 also holds for any Jacobian J_t, since the residual r_t is defined as the remainder, so two models can have the same one step error but

different propagation. Also the higher one step error of SALT is expected, because SALT is trained with rollout loss and LeWM is trained with one step loss. I think that this part needs more explanation on what is novelty in this analysis beyond the definition of the rollout error.

  1. The paper uses the comparison (4.28x for SALT vs 8.74x for LeWM) as evidence of milder propagation. However, the reproduced LeWM uses a three-frame history but the Prop. 1 is stated for a model that uses one latent state. While LeWM still uses three frames to predict, but for the propagation operator the paper only computes the Jacobian of the most recent frame and ignores the history frames (Figure 2c ,Appendix C.1, line 839-841). So LeWM's operator in Fig. 2c is only partial, while SALT's operator is complete.
Questions for authors

See Weaknesses

Reviewer_d2xKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the mismatch between one-step prediction errors and the errors that occur in recursive rollouts during planning with latent world models. The authors decompose errors during planning into the errors introduced at each step and the propagation of errors from prior steps. They then show that state-affine transitions have state-independent Jacobians and avoid nonlinear propagation residuals during planning. they then introduce SALT, which is an action-conditioned model with state-affine latent dynamics and is trained with rollout supervision. They demonstrate that SALT outperforms Le-WM at a fraction of the parameter count. They compare the models further, focusing in particular on comparing the error propagation between the two models as well as showing that SALT results in fewer failures in the OGBench-Cube environment. They further show performance gains from SALT on long horizon tasks as well as tasks involving moving two cubes.

Critical strengths

The paper provides evidence of strong empirical improvements compared to prior work using a much smaller and faster model. Additionally, the collection of experiments are interesting, including ablations for rollout training, long horizon experiments, and experiments moving more than one cube.

Outside of the empirical results, I also found the motivation and theoretical perspective to be clear, interesting, and well-motivated.

Critical weaknesses

There are some issues with Table 1 in the paper. First, the bolding seems to be incorrect. The PushT column has SALT bolded, but the highest success rate is SD-JEPA. TwoRoom is also incorrectly bolded, as the highest success rate is SMWM. Table 1 should also have all models evaluated using matched procedures as much as possible, rather than comparing models with entirely different procedures (noted with the dagger symbol). This makes it difficult to compare the performance of different models. Also, there are a number of discrepancies between the results for Le-WM here and the results reported in the original Le-WM paper. In the Le-WM paper, for example, PushT has 96% success rate, versus 87.33% here and Cube has 74% success rate, versus 67.33% here. The authors should investigate and explain the source of these discrepancies.

The mechanism for the performance improvement is also not identified precisely, and additional ablations are needed to draw a complete conclusion. The authors ablation the choice of Le-WM vs. SALT and one-step vs. rollout training objectives, but there are other differences that also should be ablated factorially. For example, Le-WM uses a three state history while SALT only uses the most recent latent and SALT has fewer parameters. There should be experiments with a three state version of SALT, a parameter matched version of Le-WM, a modification of SALT but with nonlinear transitions rather than affine, and a modification of SALT with state dependent transitions. These should each be run with one-step and rollout training.

The discussion of error propagation for Le-WM is also incomplete, although the authors note it in the text. The authors compute Jacobians using only the most recent latent state rather than the full history. The authors should run this comparison using the full history in addition, to draw an accurate conclusion.

Finally, I find the failure anatomy measures to be unintuitive and potentially confounded. First, the threshold seems a data dependent, and the authors do not demonstrate its robustness (or some intuition for why 5 is a reasonable cutoff in general). Additionally, a large ratio could simply be caused by a small denominator (mean elite cost in the final CEM iteration) or by changes in the difficulty of the planning problem in the replanning step.

Questions for authors

Why is training speed for SALT similar to that of Le-WM in spite of how much smaller SALT is?

What is the source of discrepancy in eval results for Le-WM compared with the results reported in the Le-WM paper?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.