acceptodds
Under review as a conference paper at ICLR 2027

Normalizing Flows are Capable World Models

Abstract

World models predict how the world evolves in response to actions and have developed along two main directions. Video world models, most of which are built on diffusion models, generate future observations with impressive quality but typically require large models and multiple denoising steps per prediction. Latent world models instead predict and plan in a learned representation space. Those based on Joint Embedding Predictive Architectures (JEPAs) learn this space jointly with the predictor through a regression objective, but require an additional anti- collapse mechanism, and their deterministic predictors cannot represent multiple possible futures. To bridge these directions, we present Normalizing-Flow World Model (NFWM), an autoregressive normalizing flow consisting of an invertible encoder and a probabilistic predictor. Both components are trained jointly by exact maximum likelihood, which trains the predictor to model a distribution over future latent states and prevents representation collapse without an additional anti-collapse mechanism. Since the encoder is invertible, NFWM can generate future observations like a video world model, while predicting in a single step and planning in its learned latent space like a latent world model. Although the encoder preserves all information, its latent coordinates are still learned, so we can shape them for planning with an auxiliary inverse-dynamics objective. Empirically, NFWM achieves the highest average success on deterministic goal-conditioned planning and outperforms JEPA-based world models in stochastic environments and on robotic manipulation, while also supporting policy learning in imagination. These results provide initial evidence that normalizing flows are capable world models, establishing them as a promising direction for connecting video and latent world models.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
world models, normalizing flows, JEPA

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_WS7jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Normalizing-Flow World Model (NFWM), a framework that aims to bridge video world models and latent world models. NFWM consists of an invertible encoder based on normalizing flows and an action-conditioned probabilistic predictor that models a Gaussian distribution over future latent states. The two components are jointly trained through exact maximum likelihood, which enables stochastic future prediction while preventing representation collapse without additional anti-collapse regularization. To make the learned latent space suitable for goal-conditioned planning, the authors further introduce an auxiliary inverse-dynamics objective that encourages action-sensitive representations. The paper evaluates NFWM on goal-conditioned planning in deterministic and stochastic environments, robotic action selection in LIBERO, and policy learning through imagined rollouts in Atari Boxing. The results show that NFWM achieves the highest average planning success across four deterministic control tasks, outperforms JEPA-based baselines in stochastic planning and robotic action selection, and achieves better returns in imagination-based policy learning.

Critical strengths
  1. The use of exact maximum likelihood provides a unified objective for jointly learning visual representations and stochastic dynamics. The invertibility of the encoder and the log-Jacobian term eliminate the need for additional anti-collapse regularization. The auxiliary inverse-dynamics objective provides a reasonable mechanism for shaping the latent geometry toward downstream planning.

  2. NFWM achieves competitive or superior performance compared with JEPA-based baselines across multiple control tasks, with particularly encouraging results in stochastic environments and imagination-based policy learning.

Critical weaknesses
  1. One of my main concerns is that the novelty of the proposed framework relative to existing latent world models, particularly RSSM-based approaches such as Dreamer, is not sufficiently established. Stochastic latent dynamics, generative representation learning, and policy learning through imagined rollouts have already been extensively studied in prior work. Moreover, incorporating inverse dynamics to encourage action-relevant representations is also not a new idea (e.g., Iso-Dream [1]).

[1] Iso-Dream: Isolating and Leveraging Noncontrollable Visual Dynamics in World Models. NeurIPS 2022.

  1. Moreover, the experimental comparisons focus primarily on deterministic JEPA-based world models (LeWM and SMWM). Since these baselines do not explicitly model stochastic future distributions, their inferior performance in stochastic environments is not entirely surprising. I believe that direct comparisons with established stochastic latent world models are essential to substantiate the claimed contributions and clarify whether the improvements arise specifically from the proposed normalizing-flow formulation rather than probabilistic dynamics modeling in general.

  2. NFWM operates on 144 × 192 latent features, whereas the JEPA baselines use compact 192-dimensional representations. The input resolutions also differ. Although the paper provides objective ablations, these differences make it difficult to determine which components are primarily responsible for the observed performance improvements.

  3. The robotic evaluation is limited to a single LIBERO task involving placing a mug on a plate. Moreover, the model is evaluated primarily as an action-sequence ranker over candidates generated by a pretrained policy, rather than through closed-loop planning or policy execution. It remains unclear whether the proposed approach generalizes to more diverse manipulation tasks, longer-horizon interactions, or more complex robotic environments.

  4. The paper does not provide a systematic comparison of inference latency, rollout throughput, memory consumption, or planning cost. Since the invertible encoder preserves high-dimensional image representations and the planning procedure evaluates many candidate trajectories, it is unclear how much computational advantage NFWM achieves in practice, particularly when scaling to higher-resolution observations or multi-camera robotic settings.

Questions for authors
  1. Could the authors report inference latency, rollout throughput, peak memory usage, and overall planning time compared with JEPA-based and diffusion-based world models under comparable hardware settings?

  2. Removing the inverse-dynamics objective reduces Push-T success from 86% to 34%, despite the representation remaining invertible and the dynamics predictor still being trained with exact likelihood. Does this indicate that likelihood-based representation learning primarily preserves visual information without learning a useful geometry for control?

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

NFWM jointly trains an invertible image encoder and a Gaussian latent predictor by exact observation likelihood. An inverse-dynamics loss shapes the latent space for goal-distance planning. The model supports observation generation, stochastic rollouts, and policy learning in imagination, with evaluations on navigation, manipulation, and Atari Boxing.

Critical strengths
  1. The likelihood formulation gives a clear way to combine an information-preserving encoder with stochastic prediction, without a separate collapse-prevention loss. This is a useful alternative to deterministic JEPA world models.

  2. NFWM reaches 84.5% average deterministic planning success and improves over both JEPA baselines on Teleport and Boxing. The LIBERO simulator-rollout diagnosis also separates prediction errors from weaknesses of the goal metric. The likelihood formulation and gains in both planning and imagination support weak acceptance.

Critical weaknesses
  1. The comparisons do not fully isolate likelihood training from representation design: Appendix D.1 uses a full patch state for NFWM but compact frame embeddings for LeWM and SMWM. A matched-state comparison would clarify the source of the gains. The inverse-dynamics ablation also makes clear that useful planning geometry still needs auxiliary supervision.

  2. Table 6 leaves 56% of decoded Teleport samples unassigned, while planning uses one rollout per candidate and omits terminal variance from the cost. This limits the strength of the uncertainty-modeling claim. The downstream results support the approach, but stronger distributional evaluation is needed before drawing broader conclusions about stochastic planning.

  3. Manipulation evidence comes from one custom LIBERO task. Success is 36.0%, and rises to only 42.0% when candidates are scored using simulator endpoints (Table 1). This suggests that the goal metric remains a bottleneck, and leaves generalization across manipulation tasks untested.

Questions for authors
  1. How much of the gain remains against a stochastic latent predictor with a comparable patch-state representation, keeping the training data and planning procedure fixed?

  2. On Teleport, how do success and action rankings change when the planner averages several rollouts and includes terminal variance in the expected squared-distance cost?

  3. Can you quantify valid landing rates and branch probabilities over more samples per conditioning state, following the unassigned cases in Table 6?

  4. On LIBERO, how do all three representations compare when candidates are scored using encoded simulator endpoints, as in NFWM's Table 1 diagnosis? This would clarify how much of the gap comes from goal geometry.

1 comment

Thanks for the rebuttal. The rebuttal addresses my main concerns. I have no further questions and will keep my original rating.

Reviewer_qv4ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

NFWM jointly trains an invertible image encoder and a probabilistic latent-transition predictor through exact conditional likelihood on noise-smoothed observations. The inverse encoder supports image generation, while the latent states support planning and policy learning. An auxiliary inverse-dynamics objective shapes the representation for goal comparison. Experiments cover deterministic planning, stochastic transitions, robotic action selection, and policy improvement through imagined experience.

Critical strengths

The distinction between preserving information and learning a useful planning geometry is well made. Invertibility addresses the former, while the inverse-dynamics objective addresses the latter. The authors test this distinction rather than assuming that an information-preserving representation must be suitable for control: removing inverse dynamics reduces Push-T success from 86% to 34% in the reported ablation.
The policy-learning results are also worth taking seriously. NFWM improves over its behavior-cloned initialization in all three independent runs, and the appendix documents head calibration, checkpoint selection, and an input-noise control. These checks rule out some straightforward alternative explanations for the improvement over the JEPA baselines.

Critical weaknesses

My main reservation is that the comparisons do not separate the benefit of a flow from that of stochastic prediction with a much larger latent state. LeWM and SMWM are deterministic, compact-state baselines; the paper acknowledges that Wan is not a matched comparison. The w/o NLL ablation also removes input noise and stochastic rollout, so it does not isolate likelihood training. A comparable stochastic, non-invertible world model would be a more informative test of the central design choice.

I am also less convinced by the evidence for faithful stochastic prediction. In the unselected Teleport audit, 28 of 50 NFWM outputs are not assigned to either valid landing region (Table 6). Producing both branches is encouraging, but this does not yet show that their conditional distribution is learned reliably. The appendix appropriately qualifies these counts; the planning improvement does not resolve the remaining question about transition fidelity.

Questions for authors
  1. How does NFWM compare with a stochastic non-invertible model under comparable data, predictive-state size, and planning budgets? Alternatively, can the likelihood ablation retain the same observation noise and rollout protocol?
  2. With more samples per conditioning state, what fraction depicts a valid branch, and how closely does the predicted branch distribution match the known 50:50 transition? Please include unassigned samples rather than conditioning the result on valid outputs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.