acceptodds
Under review as a conference paper at ICLR 2027

H-Lens: Extracting Phase-Space from World Model Representations

Abstract

World models hold great promise in learning representations for planning, but the dynamical structure hidden in their latent space is opaque. This also bleeds into the problem of evaluating whether the model's predictions have diverged away from reality. We introduce H-Lens, a framework for extracting phase-space dynamics from a world model. Using Hamiltonian Neural Networks (HNN), we extract coordinate and momentum-like variables out of latent states predicted by a world model. We use the learned HNN, to then extract a vector field which can be used to interpret learned dynamics. We further define a phase space error measure that predicts the rollout degradation entirely within the obtained vector field without requiring access to any ground truth states. We evaluate H-Lens on Dreamer, LeWorldModel and DINO-WM on the Pendulum, Mountain Car, and Lunar Lander environments trained on pixel observations. Models are trained and evaluated under default and dynamics-shifted environment parameters. The resulting phase space visualizations and rollout error in the vector field provide a way to inspect the dynamics learned by world models and predict their rollout degradation.

Primary area:
interpretability and explainable AI
Keywords:
world models, interpretability, physics informed networks, hamiltonian neural networks, representation learning

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

World models are neural networks that learn to predict how an environment evolves, usually inside a compressed internal representation, and they are increasingly used for planning. What dynamics they have actually learned is opaque, and it is hard to tell when their multi-step predictions have drifted away from reality. Most prior work probes these representations for individual physical quantities such as position or velocity. This paper goes further and tries to extract an explicit, interpretable dynamical system from a trained world model.

The proposed method, H-Lens, works on a frozen world model. It rolls the model out with no actions and trains a small network that compresses each internal state into a few position-like and velocity-like variables. A Hamiltonian neural network is trained alongside it: this is a network that learns an energy-like function and, assuming energy is roughly conserved, uses that function to determine how the variables evolve. Plotting the resulting flow gives a picture of the learned dynamics. The accumulated disagreement between this flow and the world model's own predicted steps is proposed as a way to detect rollout degradation without ground truth. The experiments cover three world models trained on low-resolution images of three classic control tasks (pendulum, mountain car, lunar lander), with both normal and altered physics. They show qualitatively plausible flow plots that change in the expected direction when the physics is altered. The error signal rises alongside the world model's prediction error in most, but not all, model–task combinations.

Critical strengths
  1. Works on any trained world model without retraining it. H-Lens only needs predicted sequences from a frozen model. The paper applies it unchanged to three world models built in very different ways (Dreamer, LeWorldModel and DINO-WM), so it is a reasonable starting point for inspecting what dynamics a world model has learned.
Critical weaknesses
  1. The first claimed contribution, recovering the dynamics a world model has learned, is not demonstrated. The evidence is plots, and the main comparison cannot show what it claims.
    • The Limitations section states that the recovered coordinates are only determined up to shifts, rescaling and sign flips. Yet Sec. 5 and Figs. 2–4 argue that the altered physics is recovered by comparing axis ranges between two separately trained Hamiltonian networks: pendulum velocity about 0.04 vs. 0.10, and mountain-car velocity 0.3 vs. 1.0 with position 0–1.25 vs. 2–7. Each network can rescale its coordinates freely, so these differences say nothing about the world models.
    • The axes are labelled "Angle (θ)", "Velocity (ẋ)" and so on, but the paper never measures how well the recovered variables match the true angle, position or velocity. The true states are available, since they are used to train the probes.
    • The method's core assumption, that energy is roughly conserved (Sec. 1), is never checked, and it does not hold for lunar-lander free fall or for the capped velocities and walls of the Gym pendulum and mountain car.
    • Even taken at face value, the plots show nothing that the paper's own nonlinear probe does not already show: the probe recovers the true state well for most models (Table 2), and every reading of the flow plots is checked against physics that is already known. The paper gives no case where H-Lens reveals something about a world model that was not known in advance.
  2. The second claimed contribution, detecting rollout drift without ground truth, is not demonstrated. The score would correlate with prediction error even if it carried no information, and no alternative signal is compared.
    • The phase-error score (Eq. 14) adds up a non-negative mismatch at every step, so it can only increase as the rollout gets longer. The world model's prediction error also tends to grow with rollout length. A positive correlation across rollout steps (Table 3, Fig. 5) is therefore expected from any quantity that increases over time, including the step number itself.
    • Where the two diverge, the score fails. For LeWM on LunarLander the prediction error falls while the score keeps rising, giving a correlation of −0.90 (Fig. 5h).
    • The score is never compared with step number, ensemble disagreement, or Dreamer's built-in gap between its predicted and observation-based states. The paper does not test whether it picks out which individual rollouts will fail, and it reports one training seed per model (Tables 4–6, 9).
  3. The paper is inconsistent about what was trained and evaluated. The training objective in the main text differs from the one in the appendix. The rollout length is given as at most 10 steps, 5 steps and 30 steps in different places (Secs. 3.1, 4.2, Fig. 5). A reader cannot tell exactly which model produced the reported results.
Questions for authors
  1. (Weakness 2: rollout-error score) Does the score carry any information beyond how long the rollout has run? Fix the rollout step and compare rollouts: does a higher per-step (not summed) phase error pick out the rollouts with larger prediction error? Does it do better than simple alternatives such as the step number itself, disagreement within an ensemble of world models, or Dreamer's built-in gap between its predicted and observation-based states? If it ranks failing rollouts well at a fixed step, my concern would largely go away. If it does no better than the step number, the second contribution does not hold.
  2. (Weakness 1: recovery of learned dynamics) On held-out data, how well do the recovered variables match the true angle or position and velocity after the best-fitting rescaling, for both normal and altered physics and across several seeds? Can the altered-physics claims be shown with a measure that does not depend on coordinate scale? One example is the oscillation period implied by the recovered flow, which should be shorter for the shorter pendulum. A close match and the expected change in period would support the first claim. Without them, the plots do not show that the learned dynamics were recovered.
  3. (Weakness 1: what the method adds) The nonlinear probe already recovers the true state well for most models (Table 2). Is there a case where H-Lens reveals something about a world model's learned dynamics that the probe misses and the known physics would not have predicted, such as a model whose dynamics are wrong in a way the flow plot exposes? Every interpretation in the paper is checked against physics that is already known. Without such a case, it is unclear that the method would help with its stated aim of understanding opaque world models.
2 comments

We thank the reviewer for the thoughtful comments. W2, Q1. We agree with the reviewer. In fact, we explicitly acknowledge that latent coordinates differ in basis. These changes do not alter the underlying dynamics. A pendulum does not physically swing faster simply because its coordinates are represented in a different basis. Eq. 14 measures cumulative phase-space inconsistency. Local vector-field error and accumulated rollout error are different quantities.

W1. Q2. Great question. H-Lens does not aim to recover exact physical coordinates. It aims to reveal meaningful dynamical structure. Appendix A.6 provides reference phase-spaces, while Appendix A.1 (updated) gives formal guarantees with regards to recovered dynamics.

W1. Q3. Probing requires ground-truth labels. It measures physical information in the latent representation, not the learned vector field. We use known systems for validation and unknown dynamics remain an exciting area for future work.

We thank the reviewer again for all the valuable feedback!

Thank you for the response. It reports no new results, and my main concerns remain.

Recovered dynamics (Weakness 1). The response agrees that the coordinates differ in basis between models. The paper's evidence for the altered physics compares velocity ranges across separately trained networks, which depends on that basis. No basis-independent measure is reported.

Drift score (Weakness 2). A running sum of non-negative terms grows with rollout length regardless of prediction error. The response does not compare the score with the step number or other simple signals.

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

The paper proposes H-Lens, a post-hoc method for interpreting the dynamics learned by frozen world models. An autoencoder maps world-model latent states into low-dimensional coordinate- and momentum-like variables, while a Hamiltonian Neural Network (HNN) learns a vector field describing their evolution. The method also defines a rollout-consistency score based on disagreement between the world model’s predicted trajectory and the learned vector field. Experiments evaluate Dreamer, LeWorldModel, and DINO-WM on Pendulum, MountainCar, and LunarLander, including variants with modified environment dynamics.

Critical strengths
  • The paper addresses a meaningful interpretability problem. Recovering a compact description of learned dynamics provides a potentially useful perspective beyond prediction error or reconstruction quality.

  • H-Lens is fitted after world-model training and does not require updating the underlying model. This makes it potentially useful for auditing existing models, although its applicability to larger foundation models remains to be demonstrated.

  • The visualizations make the analysis concrete. Phase portraits, overlaid trajectories, and comparisons with expected fields illustrate both successful and unsuccessful cases.

  • The paper explicitly acknowledges important limitations, including the ambiguity of the recovered coordinates and the distinction between internal consistency and ground-truth prediction accuracy.

Critical weaknesses
  • The novelty relative to prior work is insufficiently established. The related-work discussion is brief and does not clearly distinguish H-Lens from existing approaches for learning Hamiltonian dynamics from high-dimensional observations. Section 2 already describes an encoder–HNN–decoder formulation closely related to the proposed pipeline. The paper should clarify what changes beyond applying this formulation to frozen world-model representations and which aspects constitute its main methodological contribution.

  • The zero-action evaluation limits the conclusions about action-conditioned world models. The main analysis uses trajectories generated with zero actions, testing passive dynamics rather than responses to the actions a planner must compare. Furthermore, the dynamics-shift experiments train separate world models and HNNs for each condition. They therefore do not establish whether a fixed H-Lens monitor can detect an unexpected change in environment dynamics after training.

  • The evaluation is limited to three simple environments and does not establish the practical benefit of H-Lens. The experiments primarily report phase-space visualizations and relationships between internal consistency and latent rollout error. An experiment showing how this analysis improves a downstream decision would strengthen the contribution. For example, H-Lens could be used to reject unreliable candidate trajectories or determine when to replan, followed by an evaluation of whether this improves control performance.

  • The paper lacks relevant baselines for phase-space recovery and trajectory monitoring. Applying H-Lens to different world models demonstrates its applicability across architectures, but does not establish an advantage over alternative analysis methods. The state probes assess whether physical variables are recoverable; they do not provide a comparable dynamical model or monitoring score. Comparisons with an otherwise matched unconstrained dynamics model, together with ablations of the Hamiltonian and coordinate-consistency losses, would help establish the value of the proposed formulation.

Questions for authors
  • Can the authors report fixed-horizon, episode-level correlations and failure-detection performance, with uncertainty estimates and a baseline that uses only rollout horizon? How should the strongly negative LeWM–LunarLander correlation be interpreted in practice?

  • How does H-Lens compare with an otherwise matched unconstrained dynamics model? Could the main results be repeated across training seeds, and could the state probes be evaluated using episode-disjoint splits?

  • Can the authors evaluate nonzero-action trajectories? In particular, what explains the MountainCar differences under zero action? In the standard continuous environment, engine power multiplies the action and therefore contributes no direct force when the action is zero. The effects of the simultaneously changed speed limit, training distribution, and independently learned coordinates should therefore be distinguished.

1 comment

We thank the reviewer for detailed and thoughtful feedback. We address the concerns below.

Q1. We would like to point out that the paper already reports the relationship between phase-space error, rollout error at the episode level and across fixed rollout horizon (Fig 5).

The LeWM on LunarLander is an interesting exception which we discuss in Section 5. LeWM rollout curve on LunarLander increases rapidly in initial steps but then decreases. This doesn't mean the world model has started making accurate predictions again after accumulating substantial error. Rather, it means the error has saturated. Put differently, once the world model's latent predictions have deviated substantially, additional rollout steps need not produce proportional error. We already discuss this in the paper and caution against interpreting phase-space error as a direct measure of ground-truth prediction accuracy. Rather, its a useful proxy.

Q2. We appreciate the suggestion to compare H-Lens against an unconstrained dynamics model. Such an experiment would isolate the contribution of the Hamiltonian inductive bias.

We believe its important to distinguish two questions here.

  1. Whether an unconstrained model approximate learned dynamics
  2. Whether Hamiltonian bias provide an interpretable phase-space representation of the learned dynamics

The primary objective of H-Lens is the latter. An unconstrained model could approximate the same vector field without necessarily providing a position and momentum-like decomposition. Hamiltonian bias is therefore central to our objective. We address the theoretical basis of Hamiltonian bias in (updated) Appendix A.1. The formal proof in Appendix A.1 also establishes the guarantees when it comes to recovering the learned dynamics and the recovered dynamics being faithful to the learned dynamics.

Regarding training seeds and episode-disjoint probe evaluation, we do not claim that the probe results generalize to unseen episodes beyond the evaluation protocol described in the paper.

Q3. Another great question! We thank the reviewer for raising this point.

The world models are trained through uniformly sampled random actions. The dynamics-shifted are trained independently. As a consequence, they learn from different action-conditioned experiences. Thus encode different dynamics. Zero actions are only introduced later when we generate trajectories for H-Lens extraction. Zero actions do not remove the representations learned during training. However, we agree that these distinctions should be made more clear.

We again thank the reviewer for the thoughtful comments. They've helped us identify places where scope needs to be made more explicit.

Reviewer_BX2FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes to use hamiltonian neural networks to extract interpretable physical dynamics from world model latent representations. It also introduces a phase-space consistency metric to estimate rollout degradation without ground-truth observations. Experiments across three world models and physics environments demonstrate its potential for dynamics interpretation and rollout evaluation.

Critical strengths
  • The paper identifies an important problem of interpreting the dynamics learned by world models and assessing the reliability of their rollouts without relying on ground-truth observations. This is a critical step toward deploying world models in real-world applications.
Critical weaknesses

My major concern is the lack of rigorous quantitative evaluation of the proposed approach. The paper primarily relies on qualitative visualizations to demonstrate the interpretability of the extracted dynamics, without quantitatively assessing their accuracy against ground-truth physical dynamics. Furthermore, the correlation between the proposed phase-space consistency metric and actual rollout errors varies significantly across models and environments. More comprehensive quantitative experiments are needed to convincingly demonstrate the effectiveness and reliability of the proposed method.

Questions for authors

I would like to see experiments demonstrating whether the proposed phase-space error can be used to identify and select reliable world model states for downstream tasks, such as planning. Comparing downstream task performance with baselines that do not apply world model state filtering would provide stronger evidence of the practical value of the proposed approach.

1 comment

We thank the reviewer for the valuable feedback. We also appreciate the reviewer’s concerns and address them below.

1. On the validity of the recovered dynamics and the theoretical guarantee.

We would like to clarify the relationship between dynamics learned by the world model and the HNN objective. The HNN is trained on latent trajectories predicted by a frozen world model. In the HNN objective, z˙i\dot z_i is tied to world model's predictions at step tt and t+1t+1.

LHNN=1N∑i=1N∣fHNN(zi)−z˙i∣22L_{HNN} = \frac{1}{N}\sum_{i=1}^{N} \\|f_{HNN}(z_i)-\dot{z}_i\\|_2^2

=1N∑i=1N∣fHNN(zi)−EncHNN(WM(zi))∣22= \frac{1}{N}\sum_{i=1}^{N} \\|f_{HNN}(z_i)-Enc_{HNN}(WM(z_i))\\|_2^2

where fHNN(z)=∇pzH^(z)−∇qzH^(z)f_{\mathrm{HNN}}(z) =\nabla_{p_z}\widehat H(z)\\ -\nabla_{q_z}\widehat H(z)

Thus, the HNN is learning a Hamiltonian representation of the dynamics induced by the world model itself. At optimum, the HNN's derivatives match the world model's transition dynamics.

To make this more rigorous, we have formalized a proof in the revised Appendix A.1 which establishes that the HNN recovers the dynamics learned by the world model, and that the recovered dynamics reflect the learned dynamics. We believe this formal result clarifies the validity claim of H-Lens. It provides a method to expose (and characterize) the dynamics represented by a world model.

2. On the quantitative evaluation and the use of phase-space error for downstream planning.

We completely agree that the usefulness of phase-space error for downstream tasks such as planning, is an important direction. However, establishing such a result requires investigating questions about whether filtering improves planning, and how it compares with alternative strategies. We consider this a substantial future work. The primary focus of this paper is to establish the validity of H-Lens as a framework for interpreting world-models. In addition to this, phase-space consistency error is a preliminary application shown as a potentially useful signal for monitoring rollout consistency. Our intention here is not to show that it is an established method to select reliable states. Though, that remains a natural extension. As we've already acknowledged in the paper that phase-space error should be interpreted as an internal consistency signal rather than a direct measure of prediction error.

We thank the reviewer again for constructive comments. They have helped us clarify the scope of our contribution and its distinction to the more ambitious downstream planning tasks that remain to be investigated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.