acceptodds
Under review as a conference paper at ICLR 2027

Physics-Reinforced Distribution Matching Distillation for Autoregressive Video Generation

Abstract

Autoregressive (AR) video diffusion distillation enables few-step streaming generation, but matching a teacher's distribution does not by itself ensure physically plausible dynamics over extended rollouts. We study how to evaluate and improve physical plausibility in distilled AR models. We first introduce PLV-Bench, an image-to-video benchmark comprising 125 real-world videos, across 16 physical phenomena and spanning at least 10 seconds, with a reference-guided, phase-level evaluation protocol. We then present PhysDMD which augments Distribution Matching Distillation with complementary supervision from a frozen latent world model. To avoid favoring trivially predictable motion, we calibrate the generated transition's prediction energy against its paired real trajectory. The resulting signal directly supervises the final deterministic denoising step and serves as a reward for group-relative optimization of intermediate stochastic transitions. On Physics-IQ-Verified, PhysDMD improves the Overall score from 25.31 to 28.10 over Vanilla-DMD while retaining four-step sampling and requiring no world-model evaluation at inference. Ablations examine the contributions of dynamics alignment, policy optimization, and training-data composition.

Primary area:
generative models
Keywords:
Video Generation, Diffusion Models, Video Model Distillation, Few-Step Generation, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_gGjJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies the physical plausibility in AR-diffusion video models with few-step generation. It creates a new benchmark , PLV-Bench, which has real reference videos, and use a VLM to conduct evaluation. It also proposes PhysDMD, a method to augment the DMD with a reward training objective using world model (V-JEPA) signal. They show that their method can perform better than vanilla-DMD on Physics-IQ-Verified benchmark, and also their own benchmark.

Critical strengths
  1. The scope of the problem (physics plausibility in video generative models) is relevant, and the efficiency for video models is also important.
  2. They provides a new benchmakr and a new method, for evaluation and improving physics plausibility in AR video generative models.
Critical weaknesses
  1. The novelty of the paper seems limited. Using V-JEPA (world model) as a signal for physics plausibility has been studied in WMReward [1], where they use it during the inference time for search and guidance. Using RL to tilt the distribution during DMD has been studied in RTDMD [2], where they show that this combination can improve the text-to-image models' performance.
  2. The contribution of the introduction of the new benchmark seems incremental. Apart from Physics-IQ and Physics-IQ-Verified, there have been other I2V physics benchmakrs that contain real references, e.g., Physion-Eval [3] and PAI-Bench [4].
  3. The presentation of the PLV-bench is not sufficiently clear. The most concerning part is the validation of the evaluation protocol of the proposed benchmakr. There is no validation of the VLM judge performance, and it is thus hard to assess the reliability of the proposed new benchmark.
  4. The usefulness of the proposed method on open-source AR video generative models is missing. The paper applies their method on their self-trained ckpt, and shows improvment. It would be more convincing to apply their method on open-source models.

[1] https://arxiv.org/abs/2601.10553

[2] https://arxiv.org/abs/2605.26108

[3] https://arxiv.org/pdf/2603.19607v1

[4] https://arxiv.org/pdf/2512.01989

Questions for authors
  1. please explain the difference and contribution of the proposed method compared to existing literature.
  2. please explain the difference and contribution of the proposed benchmarks compared to existing literature.
  3. Do you have human evaluations on the reliability of the VLM judge?
  4. Is it possible to apply your proposed method to open sourced models? If so, how would be the results?
  5. what is the plus minus in the tables? Are they standard deviation or confidence internal or others?
  6. It seems that PhysDMD can recover the teacher model's physics performance, but cannot surpass it, as shown in ablation. Is it expected?
  7. Please double check the citation of Physics-IQ-Verified. It seems that the authors cited the original version but not the verified version.
5 comments
AuthorsOpenReview ↗

Response to Reviewer gGjJ [Part 1]

We sincerely thank the reviewer for their constructive feedback and for recognizing the importance of physical plausibility in efficient video generation. In response, we have clarified our methodological and benchmark contributions, strengthened PLV-Bench validation through human evaluation and additional analyses, and demonstrated PhysDMD's generalizability on an open-source AR video model. We have also improved our experimental reporting and citations. Detailed responses are provided below.

W1 & Q1: Contributions and Differences from Existing Methods

A1: We thank the reviewer for highlighting WMReward [1] and RTDMD [2]. PhysDMD fundamentally differs from these works in its motivation, supervision formulation, and target generation setting:

  • Motivation: Physical Plausibility Degradation. Our experiments reveal a performance gap between the multi-step AR teacher (28.22) and Vanilla-DMD (25.31) on the Physics-IQ-Verified benchmark, which continued DMD training alone fails to recover (25.99). This motivates incorporating physics-aware supervision into few-step AR distillation.
  • Versus WMReward (Training vs. Inference): WMReward employs latent world-model rewards for inference-time search and steering. In contrast, PhysDMD embeds world-model supervision directly into the distillation phase, completely bypassing inference-time overhead. Specifically, rather than directly minimizing prediction energy (which risks favoring trivially predictable motion), we introduce reference-calibrated alignment that evaluates generated transitions against paired real trajectories.
  • Versus RTDMD (AR Video vs. Image Generation): RTDMD investigates reward-guided DMD for image generation, including policy optimization of stochastic transitions and direct reward backpropagation through the final deterministic step. Building upon these optimization principles, PhysDMD introduces boundary-centered, reference-calibrated world-model supervision tailored to AR video generation, where historical errors can affect subsequent physical evolution.

Results and Core Contribution: PhysDMD improves the Physics-IQ-Verified Overall score from 25.31 to 28.10, approaching the multi-step teacher (28.22) while retaining four-step inference. On PLV-Bench, Physical Plausibility improves from 3.64 to 3.77 over Vanilla-DMD. Our core contribution is the design and integration of reference-calibrated dynamics supervision tailored to few-step autoregressive video distillation, which integrates world-model conditioning with direct dynamics alignment and reward-guided optimization across different denoising stages.

Our empirical analyses further motivate two key design choices:

1. Reference-Calibrated Reward: Directly minimizing prediction energy can favor static, easily predictable motion. To investigate this issue, we evaluated 1,000 real-video clips, comparing continuations slowed to 75% of their original speed against frozen counterparts.

Scoring RulePrefers Motion-Preserving / Normal-Speed Continuation ↑
Raw V-JEPA2 Prediction Energy32.2%
Relative Huber + Shuffled Reference57.0%
Relative Huber + Paired Real Reference (Ours)84.9%

Our symmetric relative-Huber formulation penalizes both excessively high and abnormally low prediction energies relative to the paired real transition. Paired calibration improves the preference rate by 27.9% over shuffled references, demonstrating the necessity of transition-specific supervision.

2. History-Decoupled Supervision for AR Generation: AR generation depends on previously generated frames, making world-model energy sensitive to historical errors. In a separate 1,000-clip diagnostic, replacing clean real history with generated history reversed the generated-versus-real energy ordering in 39.7% of cases.

To address this, PhysDMD decouples generation conditioning from reward evaluation: the generator is trained with augmented histories (i.e., real-video history features with augmentations) to improve robustness to imperfect rollouts, while the world model always evaluates generated and real continuations against the identical clean history. This calibrated signal supports two complementary objectives: Direct Dynamics Alignment supervises the final denoising prediction, while Reward Optimization guides intermediate stochastic transitions through GRPO. Both operate without requiring full-history self-rollouts.

Revision: We clarified our differences in Section 5.3, and expanded the supervision details in Section 3.2 and Appendices C.1–C.2.

References:
[1] Yuan, Jianhao, et al. "Inference-time physics alignment of video generative models with latent world models." arXiv (2026).
[2] Huang, Yushi, et al. "Reinforcing Few-step Generators via Reward-Tilted Distribution Matching." arXiv (2026).

AuthorsOpenReview ↗

Response to Reviewer gGjJ [Part 2]

Q2: Differences Between PLV-Bench and Existing Physics Benchmarks

A2: We thank the reviewer for highlighting Physion-Eval [3] and PAI-Bench [4]. While these benchmarks utilize real-world references, PLV-Bench fundamentally differs in its evaluation objective and how these references are used to assess sustained physical evolution.

Differences from Existing Benchmarks: Physion-Eval identifies temporally localized violations via expert annotation, and PAI-Bench broadly assesses physical perception. In contrast, PLV-Bench explicitly targets sustained physical evolution in autoregressive (AR) video generation, tackling the compounding errors and omitted intermediate stages inherent to extended long-horizon rollouts.

Methodological Design of PLV-Bench:

  1. Long-Horizon Physical Processes: We systematically curate 125 real-world videos spanning at least 10 seconds across 16 physical phenomena. We employ rigorous automated filtering and human verification to anchor multi-stage physical events.
  2. Reference-Guided Phase-Level Evaluation: Rather than directly comparing generated videos with reference frames, we construct reference-specific rubrics describing core events, ordered physical phases, and causal constraints. These rubrics are frozen before evaluating generated candidates, allowing plausible variations in timing and trajectories.
  3. Decoupled Assessment Dimensions: We independently assess Event Completion, Physical Plausibility, Temporal Evolution, and Appearance Consistency, strictly distinguishing true structural physics failures from superficial appearance degradation or incomplete progressions.

Qualitative Examples: Figure 3 visualizes the phase-level evaluation of ordered physical processes. Additionally, Appendices E.1 and F detail further reference-guided scoring examples (Figs. 9–11) and common generation failures (Figs. 16–18), respectively.

Revision: We clarified these distinctions (App. A.1), expanded the protocol (Sec. 2.3), and added qualitative examples (Apps. E.1, F).

W3 & Q3: Reliability and Human Validation of VLM-Based Physical Plausibility Evaluation

A3: We thank the reviewer for their feedback on the reliability of our VLM-based evaluation. To address this, we detail our reference-guided protocol and provide new evidence through evaluator repeatability and human validation.

Evaluation Protocol: PLV-Bench employs a two-stage process. First, the VLM constructs a frozen, reference-specific rubric (capturing core events, ordered phases, and causal constraints) strictly from the real video. Anonymized candidates are then evaluated independently via phase-level alignment rather than rigid frame-wise matching, accommodating plausible variations in timing and trajectories.

Evaluator Repeatability: Repeated GPT-5.6-Sol evaluations on 50 randomly sampled PLV-Bench cases yielded exact score agreements of 82% to 88% across all four dimensions, demonstrating high consistency.

Human–VLM Agreement: We further conducted a human study involving 11 participants and 60 generated videos spanning all 16 physical categories. Before annotation, all participants were familiarized with the evaluation guidelines, dimension definitions, and five-point scoring criteria. Each participant evaluated 40 videos, yielding 1,320 dimension-level judgments.

Scoring DimensionExact Match ↑1-Point Difference
Event Completion80.9%13.0%
Physical Plausibility77.7%11.6%
Temporal Evolution83.4%10.0%
Overall80.7%11.5%

Results: The VLM achieves 80.7% exact agreement with human assessments.

Revision: We expanded the evaluation protocol (Sec. 2.3, App. B.5), added human validation results (Sec. 4.6, Table 5), and provided qualitative examples (App. E.1, Figs. 9–11).

Reference:
[3] Zhang, Qin, et al. "Physion-eval: Evaluating physical realism in generated video via human reasoning." arXiv preprint arXiv:2603.19607 (2026).
[4] Zhou, Fengzhe, et al. "Pai-bench: A comprehensive benchmark for physical ai." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

AuthorsOpenReview ↗

Response to Reviewer gGjJ [Part 3]

W4 & Q4: Generalization to Open-Source Autoregressive Video Models

A4: We thank the reviewer for suggesting validation on existing open-source AR video models. Our primary PhysDMD model is built upon the open-source Wan2.1-14B backbone, which we causally adapt and distill into a four-step AR generator before physics-reinforced post-training.

Generalization to Causal Forcing++: To further validate the transferability of our framework, we applied PhysDMD directly to the pretrained Causal Forcing++, an independently developed open-source 1.3B AR video model with two-step inference. We preserved its original architecture and sampling configuration, applying only additional physics-reinforced post-training. All variants were evaluated on the Physics-IQ-Verified benchmark using four independent random seeds, with results reported as mean ± standard deviation.

MethodNFEST-IoU ↑Sp-IoU ↑wSp-IoU ↑MSE-based ↑Overall ↑
Causal Forcing++223.19 ± 1.6132.81 ± 0.6818.48 ± 0.3016.58 ± 0.3122.76 ± 0.35
+ PhysDMD (50 steps)223.60 ± 1.5232.75 ± 0.7118.29 ± 0.5116.88 ± 0.2822.88 ± 0.30
+ PhysDMD (100 steps)223.50 ± 1.4933.02 ± 0.9518.41 ± 0.5817.13 ± 0.3923.02 ± 0.62
+ PhysDMD (150 steps)223.53 ± 1.0833.22 ± 0.8118.54 ± 0.3917.63 ± 0.4523.23 ± 0.31

Empirical Results: PhysDMD progressively improves the Overall score from 22.76 to 23.23 within 150 additional training steps in Physics-IQ-Verified Benchmark, yielding improvements across all four metrics at the final checkpoint. Importantly, these gains are achieved while retaining the original two-step inference budget, introducing zero inference-time world-model overhead.

Revision: We added these results (Sec. 4.5, Table 4) and detailed the training configurations and intermediate checkpoints (App. D.7, Table 11).

Q5: Clarification of Statistical Reporting (Mean ± Standard Deviation)

A5: Thank you for pointing this out. Following the Physics-IQ evaluation protocol, we perform inference using four different random seeds and report the results as mean ± standard deviation. To avoid ambiguity, we have clarified this notation in the revised table caption.

Q6: Is It Expected That PhysDMD Approaches but Does Not Surpass the Teacher?

A6: We thank the reviewer for this insightful question. As stated in our Problem Formulation (Line 133), our primary objective is to recover the physical plausibility degraded during few-step AR video distillation while preserving inference efficiency, rather than strictly surpassing the teacher's generation quality.

  • Theoretical Justification: Standard DMD minimizes a reverse KL divergence [5]. This mode-seeking behavior inherently bounds student capacity [6], and distribution matching alone fails to explicitly enforce physical plausibility.
  • Empirical Restoration: As shown in Table 2, Vanilla-DMD drops significantly behind the 100-NFE AR teacher (25.31 vs. 28.22 Overall), a performance gap that continued standard DMD training fails to resolve (25.99). In contrast, PhysDMD effectively restores these physical dynamics, achieving 28.10 Overall in just 4 NFE to closely approach the teacher.
  • Fine-Grained Improvements: Notably, PhysDMD actually outperforms the teacher on specific fine-grained metrics, including mean Sp-IoU (31.13 vs. 30.09) and MSE-based scores (32.24 vs. 31.54).

Our conclusion in Sec. 6 is therefore that PhysDMD approaches the multi-step baseline while retaining four-step sampling and requiring no inference-time world model.

7. Correction of the Physics-IQ-Verified Citation.

Thank you for pointing this out. We have revised the paper accordingly and have updated the reference.

We would like to sincerely thank you once again for your time and effort in reviewing our paper.

Reference:
[5] Yin, Tianwei, et al. "One-step diffusion with distribution matching distillation." CVPR 2024.
[6] Wu, Tianhe, et al. "Diversity-preserved distribution matching distillation for fast visual synthesis." ICML 2026.

Reviewer_7yxxOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds V-JEPA2-based supervision to autoregressive video distribution matching distillation. It matches prediction energies between generated and real transitions and uses the same signal for direct alignment and GRPO.

Critical strengths

The method improves the reported benchmarks while retaining four-step inference. The comparison between alignment-only training and the full method also helps assess the effect of adding GRPO.

Critical weaknesses

The definition of the V-JEPA2 prediction energy needs clarification. Since V-JEPA2 uses masked prediction, the predictor and masking configuration determine what this energy measures. In particular, it is unclear whether the predictor sees only the history or also parts of the generated target chunk.

My main experimental concern is the lack of a direct feature regression baseline. DMD already uses perceptual regression in its original formulation, so it is important to establish what energy matching adds beyond a V-JEPA2 feature loss. Comparing feature regression and energy matching, both with and without GRPO, would address this question. The existing alignment-only comparison does not resolve it.

The experiments also rely on V-JEPA2 alone. A comparison with another suitable encoder/predictor would help determine whether the result depends on this particular model. A human evaluation of physical plausibility would be a useful addition.

Questions for authors
  • Which V-JEPA2 predictor is used, and how are the visible and target regions masked?

  • How does direct V-JEPA2 feature regression compare with energy matching, both with and without GRPO?

  • Do the improvements hold with another encoder/predictor?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7yxx [Part 1]

We sincerely thank Reviewer 7yxx for constructive feedback. We appreciate the acknowledgment of PhysDMD's performance improvements while maintaining four-step inference, as well as the recognition of our ablation studies highlighting the additional benefits of GRPO over dynamics alignment.

Clarification of Our Contribution: PhysDMD is specifically designed to tackle physical plausibility degradation in distilled few-step autoregressive (AR) video models, where conventional distribution matching fails to explicitly enforce physically realistic dynamics. To overcome this limitation, we introduce reference-calibrated world-model supervision, which aligns the prediction energy of generated transitions against paired real trajectories. Unlike direct energy minimization, which risks converging toward static or overly predictable motion, our formulation establishes a critical anchor that penalizes energy deviations in both directions relative to real physical evolution. We integrate this supervisory signal into continued DMD training via direct alignment and reward-guided optimization, while strictly preserving few-step efficiency and entirely bypassing inference-time world-model evaluation.

Q1: Clarification of V-JEPA2 Prediction Energy Configuration

A1: We thank the reviewer for requesting clarification on the V-JEPA2 prediction energy computation. We outline the key implementation details below:

  • Model Configuration: We employ a frozen V-JEPA2 ViT-L encoder-predictor (vjepa2-vitl-fpc64-256). At each autoregressive chunk boundary, we construct a 64-frame window comprising 32 ground-truth history frames and 32 continuation frames, sampled at the source frame rate.
  • Masking Strategy: A temporal block mask completely obscures the continuation, leaving solely the ground-truth history visible to the predictor. Consequently, the predictor observes zero generated frames. A separate full-clip encoder pass extracts the corresponding target features.
  • Prediction Energy: The energy is formulated as the mean cosine distance between predicted and target features across all masked tokens, with per-token layer normalization applied to the target features: \frac{1}{|\mathcal{M}|} \sum_{i \in \mathcal{M}} \left(1-\cos(\hat{z}_i,z_i)\right).$$
  • Reference Calibration: Generated and real continuations are evaluated under strictly identical history conditioning and masking. Their prediction energies are compared via our symmetric relative-Huber objective, with the real-reference energy detached from gradient computation. The V-JEPA2 model remains completely frozen during training and is discarded at inference.

Revision: We expanded Section 3.2 (Eq. 2) and Appendices C.1 and C.2 to explicitly detail this implementation pipeline.

AuthorsOpenReview ↗

Response to Reviewer 7yxx [Part 2]

Q2: Comparison Between Direct Feature Regression and Reference-Calibrated Energy Matching

A2: We thank the reviewer for suggesting a comparison with direct V-JEPA2 feature regression. To address this, we introduce a Feature Regression baseline, evaluating both alignment objectives with and without Reward Optimization on the Physics-IQ-Verified benchmark, under an identical Wan2.1-14B four-step inference configuration.

Feature Regression Baseline: Given identical history conditioning, we extract continuation features from generated and real videos via the frozen V-JEPA2 encoder and minimize their cosine distance, entirely bypassing masked prediction and energy calibration. When combined with Reward Optimization, this feature regression replaces our terminal alignment objective, while intermediate stochastic transitions retain the reference-calibrated world-model reward. Implementation details are provided in Appendix D.5.

MethodNFEST-IoU ↑Sp-IoU ↑wSp-IoU ↑MSE-based ↑Overall ↑
Without Reward Optimization
Vanilla-DMD423.01 ± 1.0932.57 ± 1.4519.06 ± 1.1926.60 ± 0.8725.31 ± 0.63
Feature Regression423.60 ± 0.9433.04 ± 1.6719.57 ± 1.1027.13 ± 1.1925.84 ± 0.66
Relative Huber (Ours)424.24 ± 0.7934.63 ± 1.3120.93 ± 0.9727.73 ± 0.7026.88 ± 0.74
With Reward Optimization
Feature Reg. + Reward Optim.425.53 ± 1.3932.47 ± 1.4019.49 ± 0.8927.63 ± 0.9826.28 ± 0.74
PhysDMD (Ours)429.42 ± 0.9931.13 ± 0.4719.61 ± 1.2432.24 ± 0.7628.10 ± 0.34

Empirical Results: Without Reward Optimization, direct feature regression achieves an Overall score of 25.84, whereas our reference-calibrated relative-Huber alignment reaches a superior 26.88. When augmented with Reward Optimization, feature regression improves to 26.28 but remains significantly below the full PhysDMD model (28.10).

These findings consistently demonstrate that direct feature matching cannot fully replicate the benefits of reference-calibrated energy alignment, even alongside reward-guided policy updates. Fundamentally, direct regression merely aligns generated representations with real target features, whereas our formulation calibrates world-model prediction energies against paired real trajectories, providing a distinctly dynamics-aware supervisory signal.

Revision: We added these comparisons to Section 4.3 and Appendix D.5 (Table 9).

Q3: Generalization Across Different World-Model Encoders and Predictors

A3: We thank the reviewer for suggesting an evaluation with an alternative world-model encoder and predictor.

Motivation for V-JEPA2 [2]: Unlike image-centric encoders (e.g., DINOv2[1]) that lack explicit future-state modeling, autoregressive video generation requires predicting subsequent frames/chunks from historical context. V-JEPA2 aligns with this pattern. Its predictive objective learns robust spatiotemporal representations (videos) to estimate future latent states, making its prediction energy an ideal dynamics-aware signal.

Generalization to V-JEPA2.1 [3]: To verify that PhysDMD is not bound to a singular architecture, we conducted a pilot experiment substituting V-JEPA2 with V-JEPA2.1. Due to limited computational resources, we ran this pilot study for only 125 H100 GPU hours. We retain the same Wan2.1-14B and four-step inference configuration for a consistent evaluation on the Physics-IQ-Verified benchmark.

MethodNFEST-IoU ↑Sp-IoU ↑wSp-IoU ↑MSE-based ↑Overall ↑
Vanilla-DMD423.01 ± 1.0932.57 ± 1.4519.06 ± 1.1926.60 ± 0.8725.31 ± 0.63
PhysDMD (V-JEPA2.1)424.83 ± 1.1132.08 ± 1.5119.49 ± 0.9128.18 ± 1.1026.15 ± 0.74

Empirical Results: The V-JEPA2.1 variant improves the Overall score from 25.31 to 26.15 over Vanilla-DMD. These results demonstrate that our framework can effectively generalize to newer iterations of predictive world models.

Revision: We added this additional experiment to Appendix D.6 (Table 10).

Reference [1] Oquab, Maxime, et al. "Dinov2: Learning robust visual features without supervision." arXiv preprint arXiv:2304.07193 (2023).
[2] Assran, Mido, et al. "V-jepa 2: Self-supervised video models enable understanding, prediction and planning." arXiv preprint arXiv:2506.09985 (2025).
[3] Mur-Labadia, Lorenzo, et al. "V-jepa 2.1: Unlocking dense features in video self-supervised learning." ECCV 2026.

AuthorsOpenReview ↗

Response to Reviewer 7yxx [Part 3]

Q4: Human Validation of Physical Plausibility Evaluation

A4: We thank the reviewer for suggesting a human validation of our physical plausibility evaluation. To address this, we conducted a human study on 60 generated videos spanning all 16 PLV-Bench categories, involving 11 participants. Prior to annotation, all evaluators were familiarized with the guidelines, dimensional definitions, and five-point scoring criteria to ensure strict protocol alignment. Each participant assessed 40 videos, yielding 1,320 independent dimension-level judgments across Event Completion, Physical Plausibility, and Temporal Evolution.

We then compared these human annotations against the GPT-5.6-Sol scores derived from our reference-guided protocol. The distribution of absolute score differences is detailed below:

Scoring DimensionExact Match ↑1-Point Difference2-Point Difference>2-Point Difference
Event Completion80.9%13.0%5.0%1.1%
Physical Plausibility77.7%11.6%9.1%1.6%
Temporal Evolution83.4%10.0%5.7%0.9%
Overall80.7%11.5%6.6%1.2%

Empirical Results: Specifically for Physical Plausibility, the VLM achieves a 77.7% exact agreement rate with human assessments, increasing to 89.3% within a 1-point tolerance.

Collectively, these findings demonstrate that the VLM provides logically sound scoring rationales. The minor deviations, predominantly manifesting as 1-point differences, likely stem from subtle variations in human understanding and the inherent subjectivity of perceiving fine-grained physical details.

Revision: We added human–VLM agreement results in Section 4.6 (Table 5) and provided qualitative examples illustrating the VLM evaluation process, scoring rationales, and justifications in Appendix E.1 (Figs. 9–11).

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

The paper proposes PhysDMD, improving physical plausibility in few-step autoregressive video generation by combining distribution matching distillation with world-model-based alignment and reinforcement learning. It also introduces PLV-Benc* for evaluating extended physical dynamics. Experiments demonstrate a promising performance.

Critical strengths
  1. Practical task: I appreciate the task of addressing physical plausibility in few-step autoregressive video generation. It is important.
  2. Well-motivated design: Using V-JEPA2 as the supervisor is interesting.
  3. Useful benchmark: PLV-Bench addresses an underexplored evaluation setting: sustained physical dynamics in few-step autoregressive video generation.
Critical weaknesses
  1. Comparison with VideoREPA[1]. VideoREPA also transfers knowledge from a pretrained video understanding model to improve physical plausibility. The paper should clarify the differences and compare against a VideoREPA-style alignment baseline.

  2. Unclear backbone setting. The paper does not clearly specify the base model or parameter count. These details should be reported for both the proposed method and the compared models to make the comparisons interpretable.

  3. Vanilla-DMD results on PLV-Bench. I am curious abot the Vanilla-DMD results on PLV-Bench. Without this baseline, it is difficult to establish how much the proposed training improves long-horizon physical dynamics.

  4. Limited evidence for improved temporal evolution. On PLV-Bench, PhysDMD scores 3.42 in temporal evolution, below LongLive-2.0 (3.50) and Helios (3.56). Given the paper’s emphasis on sustained dynamics, the authors should analyze this gap.

References

[1] Zhang, Xiangdong, et al. "Videorepa: Learning physics for video generation through relational alignment with foundation models." Advances in Neural Information Processing Systems 38 (2026): 122647-122676.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer vpde [Part 1]

We sincerely thank Reviewer vpde for their positive feedback and for recognizing the importance of improving physical plausibility in few-step autoregressive video generation. We greatly appreciate your acknowledgment of our well-motivated framework, particularly the use of a world-model as a dynamics-aware supervisor to guide video model distillation. We also value your recognition of PLV-Bench as a useful benchmark for assessing sustained physical dynamics in long-horizon video generation. Below, we provide additional analyses and clarifications to address your concerns.

Q1: Comparison with VideoREPA and Distillation-Specific Physics Supervision

A1: We thank the reviewer for highlighting VideoREPA [1], which shares a similar motivation of leveraging pretrained video understanding models to improve physical plausibility. However, PhysDMD fundamentally differs from VideoREPA in its training paradigm, supervision mechanism, and target generation setting:

  • Training Paradigm: VideoREPA fine-tunes pretrained text-to-video diffusion models by incorporating Token Relation Distillation (TRD) into conventional diffusion training. In contrast, PhysDMD addresses physical plausibility degradation during few-step autoregressive (AR) video distillation, where distribution matching alone often preserves or amplifies physically implausible dynamics.
  • Supervision Mechanism: VideoREPA transfers physical knowledge by aligning spatial and temporal token relations between pretrained understanding models and intermediate generator representations. PhysDMD instead introduces a reference-calibrated world-model prediction energy, matching generated transitions against paired real trajectories. This signal directly supervises the final deterministic prediction and guides intermediate stochastic transitions through reward optimization.
  • Target Setting: VideoREPA investigates physical commonsense in text-to-video diffusion. PhysDMD specifically targets few-step AR generation, emphasizing sustained physical plausibility under extended rollouts, evaluated via Physics-IQ-Verified and PLV-Bench.

Additional Feature Regression Ablation: Given the distinct training objectives of VideoREPA and our few-step AR distillation framework, we provide a closely related feature-regression baseline within our limited computational budget. Specifically, we conducted a controlled comparison using direct V-JEPA2 feature regression to examine whether our improvements arise merely from incorporating pretrained video representations. This baseline aligns the continuation features of generated and real videos under identical ground-truth history conditioning, without using world-model prediction energy. All variants share the same Wan2.1-14B with four-step inference, and reward optimization is excluded to strictly isolate the alignment objectives.

MethodNFEST-IoU ↑Sp-IoU ↑wSp-IoU ↑MSE-based ↑Overall ↑
Vanilla-DMD423.01 ± 1.0932.57 ± 1.4519.06 ± 1.1926.60 ± 0.8725.31 ± 0.63
Feature Regression423.60 ± 0.9433.04 ± 1.6719.57 ± 1.1027.13 ± 1.1925.84 ± 0.66
Relative Huber (Ours)424.24 ± 0.7934.63 ± 1.3120.93 ± 0.9727.73 ± 0.7026.88 ± 0.74

Results and Takeaway: Empirically, direct feature regression yields only marginal gains over Vanilla-DMD (25.84 vs. 25.31). In contrast, our reference-calibrated energy alignment achieves a superior Overall score of 26.88.

While this baseline does not perfectly replicate VideoREPA's internal TRD objective due to the distinct architectural formulations and computational constraints of adapting TRD to our 14B AR backbone, it effectively controls for representation alignment. These findings confirm that our core contribution resides in reference-calibrated physical dynamics supervision tailored for AR distillation, rather than naive feature matching.

Revision: We added this ablation to Section 4.3 (Table 2) and Appendix D.5.

Reference: [1] Zhang, Xiangdong, et al. "Videorepa: Learning physics for video generation through relational alignment with foundation models." NeurIPS 2026.

AuthorsOpenReview ↗

Response to Reviewer vpde [Part 2]

Q2: Clarification of Backbone Architecture and Model Parameters

A2: We thank the reviewer for pointing out the insufficient description of our backbone configuration. We clarify that PhysDMD is built upon the open-source WanVideo2.1-14B model [2]. Following the autoregressive distillation paradigm, our training pipeline proceeds in the following stages:

  • Causal Adaptation: We first causally adapt the pretrained WanVideo2.1-14B using a high-quality video corpus curated from public datasets (including OpenVid-1M [3]) via automated image and video quality filtering. This adapted model subsequently serves as the teacher.
  • Few-Step Distillation: We initialize and train the distilled student using OpenVidHD-derived data. The resulting student retains the 14B backbone and performs four denoising steps per autoregressive chunk.
  • Physics-Reinforced Post-Training: During this phase, our proposed method employs the frozen encoder and predictor of the pretrained V-JEPA2 ViT-L [4] to provide dynamics-aware supervision.

Revision: We clarified these backbone and training configurations in Section 4.1 and added Appendix D.1 (Table 7), which reports the generator parameter counts for all compared models and ablation variants.

Q3: Same-Backbone Baselines on PLV-Bench

A3: We thank the reviewer for suggesting the inclusion of same-backbone baselines on PLV-Bench. To directly assess whether our physics-reinforced training improves long-horizon physical dynamics beyond conventional distillation, we additionally evaluate Vanilla-DMD and Matched-step DMD on the complete PLV-Bench dataset of 125 videos. All three variants share the same Wan2.1-14B backbone and four-step inference configuration, and are evaluated under identical generation settings using the GPT-5.6-Sol reference-guided protocol.

MethodNEvent Comp. ↑Phys. Plaus. ↑Temp. Evol. ↑App. Cons. ↑
Vanilla-DMD1253.723.643.414.21
Matched-step DMD1253.663.633.404.26
PhysDMD (Ours)1253.763.773.424.34

Results: PhysDMD achieves the highest mean scores among the three variants across all four dimensions. In particular, Physical Plausibility improves from 3.64 (Vanilla-DMD) and 3.63 (Matched-step DMD) to 3.77 (PhysDMD), corresponding to gains of 0.13 and 0.14 points, respectively. Appearance Consistency also improves from 4.21 to 4.34 over Vanilla-DMD, while gains in Event Completion and Temporal Evolution remain modest. These results provide additional evidence that our physics-reinforced training improves long-horizon physical plausibility beyond standard DMD and continued DMD training, rather than merely benefiting from additional optimization steps.

Revision: We added these results to Section 4.3 (Table 3).

Reference
[2] Wan, Team, et al. "Wan: Open and advanced large-scale video generative models." arXiv preprint arXiv:2503.20314 (2025).
[3] Nan, Kepan, et al. "Openvid-1m: A large-scale high-quality dataset for text-to-video generation." ICLR 2025.
[4] Assran, Mido, et al. "V-jepa 2: Self-supervised video models enable understanding, prediction and planning." arXiv preprint arXiv:2506.09985 (2025).

AuthorsOpenReview ↗

Response to Reviewer vpde [Part 3]

Q4: Analysis and Validation of Temporal Evolution

A4: We thank the reviewer for highlighting the performance gap in Temporal Evolution. Empirically, PhysDMD (3.42) scores below LongLive-2.0 (3.50) and Helios (3.56). Our additional same-backbone evaluation (Q3) further reveals that while PhysDMD substantially enhances Physical Plausibility (3.64 to 3.77), the gains in Temporal Evolution remain modest (3.41 to 3.42).

Qualitative Analysis: This limitation is particularly evident in slowly evolving physical phenomena, such as melting, where real-world reference videos are often time-lapsed (accelerated) to capture the complete process within a short duration. Consequently, generated videos may reproduce plausible early-stage transitions but fail to reach the terminal states observed in the reference. For instance, in the ice-melting example (Fig. 18), the wet region progressively expands in the reference and LongLive-2.0, whereas it evolves relatively slowly in PhysDMD and Helios. Similarly, in the gel-melting example (Fig. 17), PhysDMD and Helios exhibit initial liquid spreading, yet the gel blocks remain upright.

Discussion and Future Directions: Following established AR distillation paradigms (e.g., CausVid [5] and Self Forcing [6]), our framework employs AR chunk-wise generation conditioned on historical context, introducing reference-calibrated physics supervision at chunk boundaries. This design effectively enhances local physical plausibility while maintaining few-step generation efficiency. However, achieving complete temporal evolution requires coherent progression across multiple physical phases, an aspect not explicitly enforced by our current localized objective. Extending physics-aware supervision to multi-chunk AR rollouts, while preserving training efficiency and stability, remains a key direction for future work.

Revision: We added this qualitative analysis to Appendix F (Figs. 17 and 18).

Reference
[5] Yin, Tianwei, et al. "From slow bidirectional to fast autoregressive video diffusion models." CVPR 2025.
[6] Huang, Xun, et al. "Self forcing: Bridging the train-test gap in autoregressive video diffusion." NeurIPS 2026.

Reviewer_8MksOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysDMD adds world-model supervision to four-step autoregressive video distillation. It compares V-JEPA2 prediction energy on generated transitions with the energy of paired real transitions. A relative Huber loss directly supervises the final deterministic denoising step. The same signal provides rewards for GRPO updates to intermediate stochastic steps.The paper also introduces PLV-Bench for reference-guided evaluation of longer physical processes. PhysDMD improves Physics-IQ-Verified scores over both Vanilla-DMD and continued DMD training, while approaching the multi-step AR baseline. On PLV-Bench, it achieves better visual consistency than the compared models and competitive physical plausibility.

Critical strengths

1.Reference calibration is a useful idea. Low prediction error can favor static or weak motion. Matching the prediction energy of a paired real transition offers a reasonable way to address this problem.
2.The experiments support the full training scheme. PhysDMD outperforms both Vanilla-DMD and continued DMD training on Physics-IQ-Verified. Its performance also approaches the multi-step AR baseline.

Critical weaknesses

1.The added value of reference calibration is not sufficiently isolated. This is the main new part of the method. However, both the hinge and relative Huber variants already use the real reference. Their comparison does not separate the benefit of calibration from that of world-model supervision. The results support the full training recipe, but leave the contribution of calibration unclear.
2. The evidence for improved long-horizon physics is limited. Table 1(b) does not include Vanilla-DMD or Matched-step DMD on PLV-Bench. It is also unclear how reliable the VLM scores are.

Questions for authors

1.What existing evidence shows the added benefit of reference calibration?
2.Please clarify how the world-model energy is computed, including the feature distance, masking scheme, and temporal window.
3.Are same-model baseline results available for PLV-Bench? What evidence supports the reliability of the VLM scores?\

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8Mks [Part 1]

We sincerely thank the reviewer 8Mks for your thoughtful feedback and for recognizing the contributions of our proposed PhysDMD framework and proposed new benchmark PLV-Bench. We particularly appreciate your acknowledgment of our reference-calibrated world-model supervision and our positive assessment of our training framework and experimental validation. Here, we address your concerns.

W1 & Q1: Contribution of Reference Calibration

A1: We thank the reviewer for suggesting we explicitly isolate the contribution of reference calibration. We agree that our initial comparison between the one-sided hinge and relative-Huber objectives primarily evaluated penalty formulations, as both already utilize paired real-trajectory energy as a calibration reference.

To address this, we introduce an Uncalibrated Energy baseline that directly minimizes V-JEPA2 prediction energy without reference subtraction or normalization. We also include a Matched-step DMD baseline to control for performance gains originating solely from continued training. For a fair comparison, all variants share the same Vanilla-DMD initialization, undergo an identical 400 training steps, and utilize Wan2.1-14B at four-step inference. We explicitly exclude reward optimization to strictly isolate the effects of alignment.

VariantNFEST-IoU ↑Sp-IoU ↑wSp-IoU ↑MSE-based ↑Overall ↑
Vanilla-DMD423.01 ± 1.0932.57 ± 1.4519.06 ± 1.1926.60 ± 0.8725.31 ± 0.63
Matched-step DMD424.80 ± 0.8632.62 ± 2.8219.43 ± 2.0827.12 ± 0.8725.99 ± 1.39
Uncalibrated Energy424.86 ± 1.1732.58 ± 1.4819.67 ± 1.1126.43 ± 1.3325.89 ± 0.90
Relative Huber (Ours)424.24 ± 0.7934.63 ± 1.3120.93 ± 0.9727.73 ± 0.7026.88 ± 0.74

Results: While uncalibrated energy minimization yields negligible improvements over Vanilla-DMD (25.89 vs. 25.31) and fails to surpass the Matched-step DMD control (25.99), our reference-calibrated relative-Huber alignment achieves a superior Overall score of 26.88 due to enhanced spatial overlap metrics.

These results support our core motivation: isolated energy minimization often leads to overly predictable dynamics. By using paired real transitions as an anchor to penalize deviations in both directions, we demonstrate the necessity of our reference-calibrated formulation over naive energy minimization.

Revision: See Section 4.3 (Table 2) for these comparisons, and Appendix D.4 for the uncalibrated baseline details.

Q2: Clarification of World Model Energy Computation

A2: We thank the reviewer for requesting further clarification on the world model energy computation. We outline the key implementation details below:

  • Model Setup: We employ a frozen V-JEPA2 ViT-L encoder predictor (vjepa2-vitl-fpc64-256). At each autoregressive chunk boundary, we construct a 64 frame temporal window comprising 32 ground truth history frames and 32 continuation frames, sampled consecutively at the source frame rate.
  • Masking Strategy: We apply a temporal block mask that hides the entire continuation, leaving only the ground truth history visible to the predictor. Consequently, the predictor observes no generated continuation frames. A separate full clip encoder pass extracts the corresponding target representations.
  • Feature Distance: We compute the prediction energy as the mean cosine distance between predicted and target features over all masked tokens, applying per token layer normalization to the target features: \frac{1}{|\mathcal{M}|} \sum_{i \in \mathcal{M}} \left(1-\cos(\hat{z}_i,z_i)\right).$$
  • Reference Calibration: Generated and real continuations are evaluated under identical history conditioning and masking configurations. Their prediction energies are compared using a symmetric relative Huber objective, with the real reference energy detached from gradient computation. V-JEPA2 remains frozen throughout training and is not required during inference.

Revision: We have expanded Section 3.2 (Eq. 2) and Appendices C.1 and C.2 to include these explicit implementation details.

AuthorsOpenReview ↗

Response to Reviewer 8Mks [Part 2]

Q3: Ablation Baselines (Vanilla DMD and Matched step DMD) on PLV-Bench

A3: We thank the reviewer for the suggestion to strengthen our long horizon evaluation. To address this, we evaluated the Vanilla DMD and Matched-step DMD baselines on all 125 PLV-Bench videos, utilizing identical generation settings and the same GPT-5.6 Sol reference-guided protocol.

MethodNEvent Comp. ↑Phys. Plaus. ↑Temp. Evol. ↑App. Cons. ↑
Vanilla DMD1253.723.643.414.21
Matched-step DMD1253.663.633.404.26
PhysDMD (Ours)1253.763.773.424.34

Results: Under these controlled settings, PhysDMD consistently achieves the highest scores across all four dimensions. The improvements are particularly notable in Physical Plausibility, where PhysDMD outperforms Vanilla DMD and Matched-step DMD by 0.13 and 0.14 points, respectively. These findings confirm that our physics reinforced formulation improves long horizon physical realism well beyond conventional DMD training.

Revision: We added these comparisons to Section 4.3 (Table 3).

Q4: Evidence supporting the reliability of the VLM scores

A4: We thank the reviewer for raising concerns about evaluator reliability. To address this, we provide additional empirical evidence demonstrating the internal consistency and human-VLM alignment of our PLV-Bench evaluation protocol.

Repeatability of VLM Scores: To evaluate the internal consistency of GPT-5.6-Sol, we conducted repeated evaluations on 50 randomly sampled PLV-Bench cases under identical settings. The evaluator demonstrates robust consistency across all four dimensions, yielding exact agreement rates between 82% and 88%.

Scoring DimensionExact Agreement ↑
Event Completion86%
Physical Plausibility82%
Temporal Evolution88%
Appearance Consistency84%

Human-VLM Agreement: We conducted a human study on 60 generated videos spanning all 16 PLV-Bench categories. Prior to annotation, all human evaluators were familiarized with the evaluation guidelines, scoring criteria, and dimensional definitions to ensure strict protocol alignment. Specifically, 11 participants assessed 40 videos each, yielding 1,320 independent dimension-level judgments across Event Completion, Physical Plausibility, and Temporal Evolution.

Against these human annotations, the VLM achieves an 80.7% exact agreement rate. The dimension-wise distribution of absolute score differences is detailed below:

Scoring DimensionSame1 Point2 Points>2 Points
Event Completion80.9%13.0%5.0%1.1%
Physical Plausibility77.7%11.6%9.1%1.6%
Temporal Evolution83.4%10.0%5.7%0.9%
Overall80.7%11.5%6.6%1.2%

Takeaway. These results provide complementary evidence of evaluator repeatability and substantial consistency with human assessments. Most VLM scores either exactly match the finalized human judgments or differ by only one point. Such small discrepancies may reflect the inherent ambiguity of fine-grained physical assessment rather than unequivocal evaluation errors. Overall, these findings support the practical utility of our reference-guided VLM evaluation protocol, while acknowledging the uncertainty inherent in subjective physical judgments.

Revision: We clarified the reference-guided evaluation protocol in Section 2.3 and Appendix B.5, added human–VLM agreement results in Section 4.6 (Table 5), and provided qualitative examples of VLM scoring outputs in Appendix E.1.

Thanks for addressing my main concerns. The new comparison with uncalibrated energy minimization supports the use of reference calibration in the direct alignment stage. The authors also clarify how the energy is computed and add both DMD baselines on PLV-Bench. The evaluator checks improve my confidence in the scores. I raise my rating from 2 to 3.

AuthorsOpenReview ↗

Thank you for your thoughtful review and recommendation

Dear Reviewer 8Mks,

We greatly appreciate your support for acceptance. We are pleased that our rebuttal has addressed your main concerns.

Best,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses to Reviewers

We sincerely thank the reviewers and AC for their constructive feedback. We appreciate their recognition of our well-motivated physics-aware AR distillation with world-model supervision (R-8Mks, R-vpde, R-gGjJ), improved few-step physical plausibility (R-8Mks, R-7yxx, R-gGjJ), and the value of PLV-Bench (R-vpde, R-gGjJ). To facilitate further AC-Reviewer discussion, we summarize below.

StrengthsR-8MksR-vpdeR-7yxxR-gGjJRevision Location
Important physics-aware AR distillation problem✓✓
Well-motivated world-model supervision✓✓
Useful reference-calibrated supervision✓
Informative alignment and GRPO ablations✓✓
Improved few-step physical plausibility✓✓✓
Strong results approaching multi-step teacher✓
Valuable long-horizon benchmark (PLV-Bench)✓✓
Clear presentation✓
Weaknesses / Questions
Insufficient calibration ablation✓Tab. 2; App. D.4
Missing feature/alignment baselines✓✓Tab. 2; App. D.5
Missing same-backbone PLV-Bench baselines✓✓Tab. 3
Insufficient human/VLM validation✓✓✓Sec. 4.6; Tab. 5
Limited temporal evolution analysis✓App. F; Figs. 16–18
Unclear world-model energy computation✓✓Sec. 3.2; App. C
Unclear backbone and experiment settings✓✓Sec. 4.1; App. D.1
Contributions vs. prior work✓✓Sec. 5.3
Limited encoder/predictor generalization✓App. D.6; Tab. 10
Limited open-source AR validation✓Sec. 4.5; Tab. 4
Rating3333

We addressed the reviewers' concerns through additional experiments, empirical validation, and method clarifications, as summarized below:

  • Reference Calibration vs. Uncalibrated Energy baseline (R-8Mks, R-gGjJ) We added an uncalibrated energy baseline, showing that reference-calibrated alignment outperforms direct energy minimization and continued DMD training. Real-video diagnostics further validate the importance of paired references.

  • Comparison with Direct Feature Regression/Alignment (R-7yxx, R-vpde): We compared V-JEPA2 feature regression with our alignment, with and without reward optimization. Our approach consistently outperforms direct feature matching.

  • Same-Backbone Evaluations on PLV-Bench(R-8Mks, R-vpde): Additional human evaluations and VLM repeatability tests demonstrate strong human–VLM agreement and consistent scoring.

  • Human Validation and Evaluator Reliability (R-8Mks, R-7yxx, R-gGjJ): We conducted additional human study and VLM repeatability tests, demonstrating strong Human–VLM agreement and consistent scoring across repeated assessments.

  • Analysis of Temporal Evolution (R-vpde): Qualitative examples reveal incomplete physical progression against time-lapsed (accelerated) references, highlighting the limitations of chunk-wise supervision for multi-stage evolution. See R-vpde (Q4, Part 3).

  • Generalization Across Different Encoders and Predictors (R-7yxx): An additional V-JEPA2.1 experiment outperforms Vanilla-DMD, providing preliminary evidence of generalization across predictive world models.

  • Generalization to Open-Source AR Video Models (R-gGjJ): Applying PhysDMD to Causal Forcing++ improves physical plausibility while preserving few-step inference efficiency.

  • World-Model Energy Computation (R-8Mks, R-7yxx): We clarified the detailed V-JEPA2 configuration and energy computation in the revised manuscript.

  • Backbone and Experimental Configurations (R-vpde): We clarified the Wan2.1-14B backbone, training settings, and baseline model parameters in the revision.

  • Method Contributions and Related-Work Comparisons (R-vpde, R-gGjJ): Motivated by physics degradation in few-step AR distillation, PhysDMD introduces reference-calibrated world-model supervision and AR-specific training designs, distinct from VideoREPA, WMReward, and RTDMD. Further details are provided in our response to R-gGjJ (Q1).

  • PLV-Bench Contributions (R-gGjJ): We clarified PLV-Bench focus on sustained physical evolution using long-horizon references (≥10s), phase-level rubrics, and decoupled evaluation dimensions, with qualitative examples.

  • Teacher Performance and Reward-Guided Distillation (R-gGjJ): We clarified our goal of recovering physical plausibility with few-step inference, including gains on certain fine-grained metrics, and discussed surpassing the teacher through reward balancing and model capacity. See more our responses to R-gGjJ (Q6 and Further Discussion).

We thank all reviewers and the Area Chair for their valuable feedback that strengthened our paper.

Best,

The Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.