acceptodds
Under review as a conference paper at ICLR 2027

Learning Hamiltonian Dynamics from Images for Long-Horizon Prediction

12 views

Abstract

Neural networks can generate visually plausible frames while failing to preserve physical consistency over time. Physics-based models introduce physical inductive biases to address this limitation, but inferring a useful state representation from images while simultaneously learning the dynamics remains challenging. We propose ViHaD (Visual Hamiltonian Dynamics), a framework that jointly learns a low-dimensional state representation from images and a scalar Hamiltonian governing its evolution. To support accurate long-horizon prediction, we penalize latent collapse and minimize prediction error over autonomous rollouts. We evaluate ViHaD on a pendulum spanning libration and rotation, a mass–spring oscillator, and a planar two-body system. Across these systems, ViHaD substantially improves visual prediction accuracy compared with PixelHNN.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Hamiltonian Dynamics, Representation Learning, Video Prediction, Physics-Informed Machine Learning

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_W6uZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ViHaD, a framework that jointly learns a low-dimensional latent state representation and Hamiltonian dynamics directly from image sequences for accurate long-horizon prediction. ViHaD incorporates higher-order centered five-point local dynamics supervision, non-collapse latent regularization, and progressive closed-loop rollout training. Evaluated on pendulum, mass-spring, and two-body systems, ViHaD significantly improves visual prediction accuracy and maintains physical consistency over extended autonomous rollouts.

Critical strengths

Optimized Training and Learning Mechanisms for Visual Hamiltonian Models: The paper proposes a series of innovative learning and training techniques (including higher-order centered five-point local dynamics supervision, non-collapse regularization constraints, and progressive closed-loop rollout training strategies). These effectively overcome long-standing issues in traditional visual Hamiltonian models (such as PixelHNN), such as latent space collapse, sensitivity to initialization, and difficulty in long-horizon training. These techniques enable visual Hamiltonian models to be trained and learned more stably and robustly, thereby achieving unprecedented long-horizon prediction accuracy in complex visual dynamical systems.

Critical weaknesses
  1. Lack of True Interpretability: Although the framework introduces Hamiltonian neural networks to model physical energy, the latent space coordinates inferred by its state encoder (such as generalized coordinates qq and conjugate momenta pp) do not correspond to true physical simulator coordinates. These latent states are merely a set of abstract mathematical representations that "happen to work" under image reconstruction losses and vector field supervision. This introduces a severe "black-box" property: users cannot directly interpret or extract true physical laws (such as absolute physical scales or constants) from these latent states, nor is it easy to safely bridge such abstract latent spaces—devoid of physical anchors—into downstream embodied intelligence control or real-world robotic interactions.

  2. Overly Simplistic Evaluation Scenarios and Lack of Real-World Video Validation: All experiments in the paper are limited to ideal synthetic datasets perfectly rendered by computers using mathematical formulas (including pendulum, mass-spring, and planar two-body systems). These scenarios feature clean backgrounds, no lighting variations, no occlusions, and strictly adhere to the assumption of friction-free conservative systems. However, the framework has never been validated on any real-world videos featuring complex background noise, lighting interference, motion blur, and frictional dissipation. This casts a huge question mark over its generalization capability and robustness when handling real-world visual noise.

Questions for authors
  1. Have the authors tested this framework on real-world videos containing background noise, lighting interference, motion blur, or frictional dissipation? When facing real-world visual noise, can the encoder EψE_\psi still reliably extract well-behaved phase-space states?

  2. ViHaD is strictly built upon conservative Hamiltonian systems (assuming total mechanical energy conservation). However, the vast majority of real-world physical processes involve friction, damping, or energy dissipation. How do the authors plan to extend this framework to non-conservative, damped, or externally controlled systems?

  3. Although the model employs a Hamiltonian structure to constrain dynamics, the inferred latent space coordinates (qq and pp) do not correspond to true physical simulator coordinates. How can the physical interpretability be verified in this case?

4 comments
AuthorsOpenReview ↗

Response to Reviewer W6uZ — Part 1/2: Interpretability and evaluation scope

We thank the reviewer for the careful reading and for recognizing the long-horizon improvements over PixelHNN.

W1. Lack of true interpretability

We agree that ViHaD does not recover simulator coordinates, absolute physical units, or physical constants, and we do not claim that it does. The goal is to learn a dynamically useful Hamiltonian latent phase space rather than an identifiable simulator coordinate system.

The learned representation nevertheless preserves meaningful physical organization. Without regime labels, the pendulum latent space separates libration and both rotation directions, with physical energy varying smoothly within each family (Figure 3). On mass–spring, circular physical trajectories map to nested, energy-ordered closed latent orbits (Figure 6). Appendix A.5 further uses one fixed post-hoc readout, trained only after ViHaD, to map pendulum latents to physical-state variables for evaluation. These results support dynamical-structure recovery, not coordinate-wise identification.

Paper reference: the method explicitly states that the learned coordinates need not match simulator coordinates in Lines 155–161. The phase-space analysis is in Appendix A.5, Lines 864–889, with the held-out/readout rollout comparison in Figure 7, Lines 946–952.

W2. No real-world video validation

We agree that the current evaluation is limited to controlled conservative systems. This is intentional: the experiments are designed to isolate latent-state learning and error accumulation during long autonomous integration under known dynamics.

PixelHNN, HGN++, COMET and Hamiltonian GAN also evaluate their image-based models on rendered simulations.

We do not claim robustness to real-world video with complex backgrounds, lighting changes, occlusion, or motion blur. These settings are outside the present scope. The paper's conclusion already identifies real-world video as a future direction. Our present contribution is therefore narrower: establishing the behavior of the proposed representation-learning and rollout-training mechanisms in controlled visual Hamiltonian systems.

Paper reference: the conclusion limits the forward-looking claim to future work on more complex dynamics and real-world video in Lines 484–490. No real-world robustness claim is made elsewhere in the current revision.

AuthorsOpenReview ↗

Response to Reviewer W6uZ — Part 2/2: Real-world video, extensions, and physical interpretation

Q1. Have you tested real-world video or visual noise?

Not yet. The current experiments use clean rendered observations with known underlying dynamics so that representation learning and rollout behavior can be evaluated without confounding visual factors.

We therefore cannot claim that the current encoder will remain reliable under real-world background noise, lighting variation, motion blur, or occlusion. Testing and adapting ViHaD for such conditions is an important next step, and the conclusion already identifies learning from real-world video as a future direction.

Paper reference: real-world video is explicitly identified as a future direction in the conclusion, Lines 488–490. This point is a scope clarification rather than a newly added experiment.

Q2. How would ViHaD extend to dissipative or controlled systems?

The present vector-field parameterization assumes conservative autonomous Hamiltonian dynamics. Extending ViHaD to damped, forced, or controlled systems would require changing the dynamics model rather than treating those effects as if they were conservative.

A natural direction is to augment or replace the conservative Hamiltonian vector field with an appropriate dissipative, port-Hamiltonian, or controlled-Hamiltonian formulation; related examples are already discussed in the paper's Related Work. The representation-learning components, higher-order local supervision, non-collapse regularization, and closed-loop rollout-training strategy could then be used with that modified dynamics parameterization. We do not claim that the current model already handles these settings; they are future extensions.

Paper reference: the Related Work already points to controlled and non-conservative Hamiltonian extensions—Symplectic ODE-Net, port-Hamiltonian networks, and dissipative Hamiltonian models—in Lines 87–90. The current method itself is explicitly introduced for conservative mechanical systems in Lines 152–153. The proposed extension remains future work and is not claimed as part of the current model.

Q3. How can physical interpretability be verified if latent q,pq,p differ from simulator coordinates?

We distinguish physical-coordinate recovery from dynamical-structure recovery. ViHaD does not guarantee that its learned q,pq,p coincide numerically with the simulator coordinates. Hamiltonian dynamics are not unique to a single coordinate representation, and the model receives no simulator-state labels that would anchor the latent coordinates to the simulator's units or scale.

What we verify is that the learned representation contains and organizes physically meaningful state information. For the pendulum, the latent space separates the three motion families and is smoothly ordered by physical energy within each family. A single fixed post-hoc readout, fitted on training-split encodings and never used during ViHaD training, maps held-out encoded and rolled-out states to coherent physical phase portraits (Appendix A.5, Figure 7). For mass–spring, encoded and autonomous trajectories form nested energy-ordered closed orbits (Figure 6).

These analyses support the claim that the latent state is dynamically meaningful; they do not establish identifiable recovery of simulator coordinates or physical constants.

Paper reference: the non-identifiability of the learned coordinates is stated in Lines 155–161. Appendix A.5 gives the phase-space evidence in Lines 864–889, and Figure 7's held-out encoded/rollout/ground-truth comparison is described in Lines 946–952.

Reviewer_W6uZOpenReview ↗

Response to rebuttal

Thank you for your explanation. However, I am not convinced that the proposed model will work beyond toy examples. In addition, as for its theoretical contribution, the physical interpretability is not clear, or at least not well proven. Thus, I maintain my initial rating.

AuthorsOpenReview ↗

Clarifying ViHaD’s dynamical interpretability

Thank you for the follow-up. We agree that the absence of real-world experiments is a limitation. We do not claim real-world robustness; extending ViHaD to real video is left as future work. Our evaluation uses controlled rendered systems, as is common for visual Hamiltonian methods, so that long-horizon rollouts can be checked against known dynamics.

To clarify the interpretability point: the paper does not present physical interpretability as a contribution, and makes no theoretical identifiability claim. Two notions should be separated.

Coordinate identifiability asks whether a latent coordinate equals the simulator's angle, momentum, scale, or physical parameters. ViHaD does not claim this: it is trained without simulator-state supervision, so its latent coordinates are not assumed to match the simulator coordinates (Lines 161, 173–174), nor do they need to. The same Hamiltonian dynamics can be written in many coordinate systems, and accurate prediction only requires that the model learn one of them consistently.

Recovery of dynamical structure asks whether the latent space arranges trajectories the way the physics does. This is what Appendix A.5 evaluates.

  • Mass–spring (Figure 6). Physically, each trajectory is a closed loop, and higher-energy loops enclose lower-energy ones. The latent space shows the same picture: one closed loop per trajectory, nested in the same energy order, stretched in shape but never crossing. Autonomous 100-step rollouts, started from a single encoded state with no further observations, closely retrace these loops. No readout is involved.
  • Pendulum (Figures 3 and 7). A pendulum can swing back and forth or spin fully around in either direction. Without any labels, the latent space places these three kinds of motion in three separate, non-overlapping regions. To view this in physical terms, we fit one readout to (q,p)(q,p) after training and apply it unchanged to both encoded observations and rollouts. Across 28 held-out trajectories and 192 steps, every swinging trajectory keeps swinging and every spinning one keeps spinning in its original direction, as in the ground truth.

Our claim is therefore dynamical-structure recovery, not coordinate-wise recovery of simulator variables or physical constants.

Reviewer_HcbtOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper tries to predict simple mechanical systems from video observations alone, without access to the ground truth physical states of the system. ViHaD uses an encoder to turn short history of past frames into low-dimensional latent sates, that is split into momentum and position. A hamiltonian network defined the dynamics of the state on the latent space. On top of PixelHNN the proposed method adds a five-point derivative target as well as loss fucntion that stop the latent from collapsing.

Critical strengths

The paper demonstrates it is able to learn a 2-dimensioanl pahgse space of a pendulum across swining/spinning in both directions and rolling out to over 100 steps using just 6 history frames. IoU and other metrics remain significantly above the baseline PixelHNN, which was originally only tested at small angles where the dynamics are much more well behaved.

Critical weaknesses

Theoretically I think the paper misses several existing works that study the latent space of dynamical systems, such as [1,2], which uses different forms of regularization to ensure that the latent space is well behaved. Furthermore the method should really be baselined against methods such as HGN/Hamiltonian GAN and non-structure preserving baselines. As it stands using just PixelHNN is simple too old of a baseline to show state of the art performance. The tested dataset are also far too simplistic to fully show case that the proposed method represents a significant jump over existing methods.

[1] Kasim, Muhammad Firmansyah, and Yi Heng Lim. "Constants of motion network." Advances in Neural Information Processing Systems 35 (2022): 25295-25305.
[2] Allen-Blanchette, Christine. "Hamiltonian gan." 6th Annual Learning for Dynamics & Control Conference. PMLR, 2024.\

Questions for authors

Q1. For the pendulum and two-body rows of Table 1, which seed and checkpoint are reported, and how were they chosen?

Q2. Similar to baseline methods can the author show energy conservation of the learned rollout? Right now there is no evidence that the proposed method actually achieves conservation results better than baselines.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Hcbt — Part 1/2: Related work, baselines, and datasets

We thank the reviewer for the detailed feedback, particularly the requests for broader baselines, missing related work, checkpoint transparency, and direct physical-energy analysis.

W1. Missing related work: COMET and Hamiltonian GAN

We agree that both works are relevant and have added them to Related Work.

COMET (Kasim & Lim, 2022) learns a vector field jointly with constants of motion and constrains the learned field to conserve them. Its experiments primarily use simulator states, with one image-based two-body experiment. Hamiltonian GAN (Allen-Blanchette, 2024) is an unconditional video generator: it samples an initial phase-space state from noise, evolves it using learned Hamiltonian dynamics, and evaluates generated video samples.

Neither work addresses our setting: conditional long-horizon prediction from images across all motion regimes of a pendulum, including libration, both rotation directions, and near-separatrix motion. We have nevertheless added both works because they are directly relevant to learned dynamical structure and visual Hamiltonian modeling.

Revision in paper: COMET is added in Lines 93–95; Hamiltonian GAN and its distinction from conditional forecasting are added in Lines 104–106 (continuing at Line 131 after the figure placement).

W2. Missing HGN/Hamiltonian GAN and non-structure-preserving baselines

We agree that PixelHNN alone was too narrow a baseline set. We added two pendulum baselines: HGN++, the causal successor of HGN, using the authors' released code; and an unconstrained vector field, a non-structure-preserving variant of ViHaD.

Model — pendulum, 800 test trajectoriesMean IoUIoU@100
PixelHNN0.07410.0424
HGN++0.76310.2118
Unconstrained vector field0.59830.5275
ViHaD, progressive0.91170.8102

For the unconstrained baseline, we replace only J∇zHθ(z)J\nabla_z H_\theta(z) with the two outputs of the same MLP used directly as the latent time derivative, while retaining the same losses, loss weights, horizon schedule, and learning rates. HGN++ stays close to ViHaD through its 60-step training window but falls to 0.2118 IoU at step 100.

HGN and Hamiltonian GAN are not like-for-like causal forecasting baselines: HGN infers its state from the whole sequence it reconstructs, while Hamiltonian GAN generates videos unconditionally rather than inferring an initial state from observed frames. The new HGN++ and unconstrained rows are included in Table 1 of the revised paper.

Revision in paper: the new HGN++ and unconstrained rows are in Lines 325–341; Figure 2(a) and its caption include both baselines in Lines 343–358; their comparative discussion is in Lines 361–369; Appendix A.6 gives implementation details in Lines 1004–1034.

W3. Datasets are too simplistic

We agree that the systems are visually simple. We chose rendered conservative systems deliberately to isolate one problem: learning from images a latent state and dynamics that remain accurate under long autonomous rollouts. The pendulum is nevertheless nontrivial dynamically because one model must represent libration, both rotation directions, and near-separatrix motion, and the two-body experiment uses an eight-dimensional latent state.

PixelHNN, HGN++, COMET, and Hamiltonian GAN also evaluate their image-based models on rendered simulations. We do not claim that the present results establish robustness to real-world video or non-conservative dynamics. The conclusion already identifies real-world video as a future direction.

Paper reference: the controlled pendulum setup is described in Lines 258–269, the two-body setup in Lines 279–291, and the conclusion explicitly identifies real-world video as future work in Lines 484–490. No real-video robustness claim is added.

AuthorsOpenReview ↗

Seed selection and physical-energy analysis

Q1. Which seed and checkpoint are reported for the pendulum and two-body rows, and how were they chosen?

Table 1 reports the best seed for every row except mass–spring ViHaD, which reports the seed closest to its four-seed mean. Seeds, checkpoints, and loss weights were selected by test-split Mean IoU for the baselines and for ViHaD alike.

For the non-progressive rows, the available-seed final-checkpoint statistics are:

System / methodSeedsReported Mean IoUAll seeds, final checkpoint — mean ± sd
Pendulum, PixelHNN50.07410.0682 ± 0.0063
Pendulum, ViHaD H=1H=160.21420.1213 ± 0.0495
Two-body, PixelHNN60.57340.4825 ± 0.0539
Two-body, ViHaD H=1H=140.72460.6304 ± 0.0832

The progressive models are single training lineages, with each stage initialized from the preceding stage's selected checkpoint. The reported pendulum lineage is seed 2 and reaches Mean IoU 0.9117 at the H=32H=32 stage; the reported two-body lineage is seed 1 and reaches 0.9438. For two-body, checkpoint/configuration comparisons were made on a 2,000-trajectory subset of the 10,000-trajectory test set. We provide these details explicitly for transparency.

Paper reference: the training schedules for the two-body progressive model are in Lines 684–690 and for pendulum progressive ViHaD in Lines 694–703. The detailed seed identities and selection procedure above are response-only and are not added to the current revised paper.

Q2. Can the authors show energy conservation of the learned rollout?

Yes. We added a quantitative physical-energy analysis for the pendulum. We use one state estimator, trained once on ground-truth training frames to map six consecutive frames to (cos⁡q,sin⁡q,p)(\cos q,\sin q,p), and apply that same estimator to every model's predicted frames. It never sees model predictions during training. From its output we compute

E^=p^2+1.5(1−cos⁡q^)\widehat E=\widehat p^2+1.5(1-\cos\widehat q)

and compare it with the trajectory's constant true physical energy.

The estimator's energy MSE on ground-truth test frames is 1.9×10−5 J21.9\times10^{-5}\,\mathrm{J}^2. The measured prediction error also includes estimator error on generated frames, so it should be interpreted as physical-energy error measured through this common estimator rather than a certified bound on the underlying latent dynamics.

ModelEnergy MSE, mean steps 1–100 (J2\mathrm{J}^2)Energy MSE at step 100 (J2\mathrm{J}^2)
PixelHNN4.834.98
Unconstrained vector field2.312.41
HGN++0.502.60
ViHaD, progressive6.1×10−36.1\times10^{-3}7.4×10−37.4\times10^{-3}

ViHaD's energy MSE remains below 9×10−3 J29\times10^{-3}\,\mathrm{J}^2 at every rollout step. HGN++ is very accurate early in the rollout but its energy MSE grows sharply after its training window, reaching 2.60 J22.60\,\mathrm{J}^2 at step 100. The revised paper includes the estimator description, energy-vs-step figure, and compact comparison table in Appendix A.7.

Revision in paper: the main-text results point to the energy analysis in Lines 367–369. Appendix A.7 describes the shared estimator and evaluation in Lines 1037–1053 and reports the comparative behavior in Lines 1055–1059. Figure 11's caption is in Lines 1101–1103, Table 12 is in Lines 1106–1112, and Figure 12's caption is in Lines 1153–1155.

I thank the authors for their rebuttal. While I think the paper is stronger with the inclusions I do not think it moves the needle to a higher score for me. I think running the proposed method on something more substantial would've shown the methods capability rather than leaving it to future work, even if the relatively old baselines did not do so.

For the AC, my evaluation is borderline, if you will, 2.5...

Reviewer_UpsPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ViHaD, a framework for learning latent state representation and neural Hamiltonian dynamics from visual observations. Building upon the PixelHNN framework, the authors introduce higher-order finite-difference supervision, latent non-collapse regularization, and progressive closed-loop rollout training. Experiments on pendulum, mass–spring, and two-body systems demonstrate substantial improvements over PixelHNN.

Critical strengths

S1. Learning physical dynamics directly from visual observations and leveraging them for physically consistent prediction is an important and meaningful research problem.

S2. The proposed three module designs are well motivated and provide practical improvements over PixelHNN.

S3. The experiments convincingly demonstrate substantial improvements over PixelHNN across three physical systems. The ablation studies and latent space visualizations further provide useful insights into the effectiveness of the proposed components.

Critical weaknesses

W1. Limited contribution. Learning latent states and Hamiltonian dynamics directly from images has already been explored by prior works such as PixelHNN and HGN. ViHaD primarily refines the training objectives and optimization strategies within this established framework, making its contributions incremental.

W2. Missing baselines. Although the paper identifies HGN and HGN++ as closely related works, the comparisons are limited to PixelHNN. The absence of these important baselines makes it difficult to assess the actual progress of ViHaD over existing methods.

W3. Unclear claim. The paper lacks a clearly articulated central claim that distinguishes it from prior work. Specifically, is the primary focus on learning well-structured latent representations, or on improving long-horizon prediction through refined training strategies?

Questions for authors

Q1. Why do the authors choose a five-point centered finite-difference scheme instead of a three-point or seven-point scheme?

Q2. Why does PixelHNN perform so poorly on the pendulum task? The paper lacks a sufficient explanation for this observation.

If the authors adequately address my concerns above, I would be willing to adjust my score.

1 comment

Thank you for the rebuttal. I acknowledge the improved prediction performance, but my concerns remain:

The paper lacks a comparative analysis of the representations learned by the baselines. It therefore remains unclear whether the proposed components indeed lead to more dynamically useful representations, as motivated.

Furthermore, it is still unclear why the proposed components, can effectively address the central motivation of learning a dynamically useful latent representation. In addition, the connections between these components are not sufficiently clear, making the method appear more like a collection of separate modules.

Overall, I acknowledge that the method works to some extent, but the above concerns remain. I therefore lean to maintain my Weak Reject recommendation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Author Response — Summary of Revisions

We thank all three reviewers for their careful reading and constructive feedback. The discussion highlighted four recurring issues: clarifying ViHaD's central claim, broadening the baseline comparison, justifying the choice of the five-point stencil, and providing additional physical evidence for long-horizon rollouts.

Our central claim is about long-horizon causal prediction from images: ViHaD learns, from images alone, a latent state and Hamiltonian dynamics that support accurate autonomous prediction over long horizons. A well-behaved latent phase space and closed-loop rollout training are the means to that goal. In the revised paper, changes are marked in blue.

We added the following evidence and revisions:

  • HGN++ baseline: on the full-regime pendulum, HGN++ reaches Mean IoU 0.7631 and IoU@100 0.2118, compared with 0.9117 and 0.8102 for progressive ViHaD. HGN++ stays close to ViHaD over its 60-step training window but degrades substantially beyond it.
  • Non-structure-preserving baseline: replacing the Hamiltonian vector field with an unconstrained vector field while retaining the ViHaD training losses and rollout schedule gives Mean IoU 0.5983 and IoU@100 0.5275.
  • Physical-energy analysis: on the pendulum, ViHaD's predicted-rollout energy MSE remains below 9×10−3J29\times10^{-3} \mathrm{J}^2 at every rollout step. The revised appendix reports the estimator, energy-vs-step curve, and comparison with PixelHNN, HGN++, and the unconstrained baseline.
  • Three-/five-/seven-point comparison: The revised paper compares derivative-target error on true pendulum states, five-seed direct-state rollout error, and four-seed image-prediction performance on mass–spring. Five-point clearly improves over three-point and achieves image-prediction performance comparable to seven-point with a narrower temporal neighborhood.
  • Related work: we added COMET (Kasim & Lim, 2022) and Hamiltonian GAN (Allen-Blanchette, 2024).
  • Seed/checkpoint details: we provide the seed and checkpoint-selection procedure for the pendulum and two-body results.

The new baseline experiments focus on the pendulum because it spans the widest range of qualitatively distinct motion regimes.

Current revised-paper locations: central claim and contributions, Lines 37–69; COMET, Lines 93–95, and Hamiltonian GAN, Lines 104–106 and 131; new pendulum baselines and rollout discussion, Lines 325–369; stencil comparison, Lines 461–479 and Appendix A.4, Lines 772–845; HGN++/unconstrained implementation details, Appendix A.6, Lines 1004–1034; energy analysis, Appendix A.7, Lines 1037–1155.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.