Pith. sign in

REVIEW 5 major objections 8 minor 52 references

The paper claims that reweighting preference pairs by semantic divergence bounds the semantic drift caused by physical-preference optimization in text-to-video models, roughly doubling physical plausibility without sacrificing prompt fideli

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 19:27 UTC pith:ISEPJJFU

load-bearing objection A real diagnosis and a sensible reweighting fix for DPO in T2V, wrapped in theoretical claims that are mostly restatements of the loss definition and experiments that lack error bars. the 5 major comments →

arxiv 2607.16947 v1 pith:ISEPJJFU submitted 2026-07-18 cs.CV

When Physical Preferences Meet Semantic Constraints: Physical and Semantic Direct Preference Optimization for Text-to-Video Generation

Siwei Meng , Yawei Luo , Shu Zhang , Ping Liu This is my paper
classification cs.CV
keywords text-to-video generationdirect preference optimizationphysical commonsensesemantic alignmentpreference alignmentconstrained optimizationsemantic driftvideo generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard direct preference optimization for physical realism in text-to-video generation assumes the physically preferred video also matches the prompt. The paper shows this assumption fails systematically: in about half of physical preference pairs, the video judged more physical is actually less aligned with the text. To fix this, the paper frames physical alignment as a constrained problem and proposes PSDPO, which reweights each preference pair by the semantic discrepancy between chosen and rejected videos. A gradient analysis shows this bounds semantic drift to a controllable residual, and a two-stage training schedule provably reduces cumulative drift.

Core claim

The central claim is that physical-semantic conflict in preference-based video alignment is not a data artifact but a structural consequence of comparing dynamics without checking prompt fidelity. PSDPO treats semantic consistency as a feasibility constraint and modulates the DPO loss per pair with a weight W(ΔS) that is 1 when the physical winner is also semantically better and decays exponentially when the physical winner is semantically worse. Proposition 3.1 bounds the gradient residual from conflicting pairs by G·E[exp(ΔS/τ)], and Corollary 3.2 shows that training first on non-conflicting pairs and only later on conflicting pairs reduces the cumulative drift bound by T_s/T_total. The me

What carries the argument

The central object is the asymmetric modulation weight W(ΔS)=1 for ΔS≥0 and exp(ΔS/τ) for ΔS<0, where ΔS is the semantic alignment score of the physically preferred video minus that of the rejected one. This weight converts the constrained problem into a single reweighted DPO objective, preserving consistent pairs unchanged while suppressing conflicting pairs in proportion to their semantic violation. The companion mechanism is the two-stage curriculum: stage one trains only on consistent pairs, stage two adds conflicting pairs with the modulation, which the paper shows reduces cumulative drift.

Load-bearing premise

The load-bearing premise is that the semantic discrepancy ΔS, computed once on videos generated by the initial supervised-fine-tuned policy, remains a valid measure of semantic constraint for the policy being trained; if those labels go stale as training changes the model, the constrained objective and its drift bound apply to the dataset rather than to the evolving distribution.

What would settle it

Recompute the semantic scores at the end of PSDPO training on fresh samples from the trained policy for the same prompts, and compare the proportion of physical-preference pairs that are semantically inverted against the proportion at initialization. If the inversion proportion does not drop (or the semantic score of the chosen videos degrades as much as under vanilla DPO), then the claimed drift bound is not binding on the deployed policy. Alternatively, ablate the weight W(ΔS) by replacing it with a constant and checking whether the semantic drift returns, which would confirm the mechanism.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The method operates entirely within the standard DPO framework, requiring no auxiliary models or additional loss terms, so physical alignment can be achieved by reweighting existing preference pairs.
  • The temperature τ interpolates between hard constraint satisfaction (τ→0) and vanilla DPO (τ→∞), giving a single scalar that controls the physical–semantic trade-off.
  • Corollary 3.2 implies that delaying the introduction of conflicting pairs by T_s training steps cuts the cumulative drift bound by T_s/T_total, so staging is not a heuristic but a provable improvement.
  • On the reported benchmarks, PSDPO doubles physical plausibility on a hard physical subset relative to the base model while keeping semantic scores stable or better.
  • The same training pipeline extends to a larger text/image-to-video backbone with stable loss and preference-accuracy curves, indicating the recipe is architecture-independent.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The analysis suggests a general recipe for any multi-objective preference alignment: if one signal is pairwise comparative and the other is per-sample, replace hard filtering with a smooth weight based on the per-sample discrepancy; the exponential form is one admissible choice among D1–D4.
  • The bound relies on ΔS being measured once on the initial policy; a natural extension is to recompute or adaptively update the semantic scores during training, which would likely tighten the residual estimate further.
  • The staged-curriculum insight may transfer to other domains with conflicting supervision, such as aligning language models where helpfulness preferences clash with safety constraints—feasible pairs first, then modulated conflicting pairs.
  • The paper's failure-case analysis suggests an untested implication: PSDPO is limited by base-model priors; an explicit prompt-augmentation pass that states physical outcomes could combine with PSDPO to address rare physics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 8 minor

Summary. The paper addresses semantic drift in preference-based physical alignment for text-to-video generation. It observes that in a large fraction of physical preference pairs, the physically preferred video is semantically less aligned with the prompt than the rejected video. To mitigate this, the authors formulate physical alignment as a constrained preference optimization problem and propose PSDPO, which reweights the DPO loss per pair with W(ΔS)=exp(ΔS/τ) for conflicting pairs, plus a two-stage training protocol that first trains on match pairs and then introduces conflict pairs. They claim a gradient-level bound (Proposition 3.1) on semantic drift and a provable reduction of cumulative drift via staging (Corollary 3.2). Experiments on VBench, VideoPhy-2, and PhyGenBench report improvements over existing preference-based methods, including a 2× gain on the VideoPhy-2 Hard subset, while maintaining VBench semantic scores.

Significance. The empirical finding that semantic inversion affects over half of physical preference pairs is valuable and actionable. The proposed reweighting is simple and operates within the standard DPO pipeline, and the ablations on τ, Ts, and the staged protocol are thorough. The paper also includes a human study validating its PCE preference construction system, an additional backbone experiment, and an honest failure-mode analysis. If the empirical results hold, the method is a practical contribution. However, the theoretical claims as written are not supported: the constraint in Eq. (2) is on fixed dataset videos rather than on the policy, and Proposition 3.1 bounds a weighted gradient norm without connecting to semantic drift of generated outputs. These issues need to be resolved before the claims of formal guarantees can be accepted.

major comments (5)
  1. [Section 3.1, Eq. (2)] The constrained optimization is ill-posed over θ. The constraint S_sem(y_w,x) ≥ S_sem(y_l,x) is evaluated on the fixed videos in D and contains no dependence on π_θ. With 51.5% of pairs violating the inequality, the feasible set is empty, so the 'constrained' formulation does not define a meaningful optimization problem. The reweighting in Eq. (4) is introduced as a relaxation, but the paper does not derive it from Eq. (2) as a Lagrangian, penalty, or projection. This breaks the claimed connection between the constrained formulation and the PSDPO objective.
  2. [Section 3.3, Proposition 3.1] The residual ε is defined as E_conf[W(ΔS)∇ℓ], and the bound ||ε|| ≤ G·E_conf[W(ΔS)] follows immediately from the triangle inequality and the bounded-gradient assumption. This is a definitional restatement of the reweighting, not an analysis of semantic drift. The proposition does not bound any semantic quantity of the policy’s outputs (e.g., a text-video alignment score on generated samples). Calling this a bound on 'semantic drift' is a misnomer; the result is a bound on the norm of a fixed-data weighted gradient. The central theoretical claim is therefore not established.
  3. [Section 3.4, Corollary 3.2] The cumulative residual reduction is a counting argument. Since ||E_direct|| ≤ T_total·B and ||E_staged|| ≤ (T_total−T_s)·B, the improvement is just the consequence of summing over fewer terms. There is no analysis of how gradients from D_match interact with later gradients from D_conf, no treatment of optimization dynamics, and no reason that early training on D_match would reduce the per-step residual bound B. The corollary does not 'provably reduce cumulative drift' in any substantive sense beyond the definition of the sum.
  4. [Section 4.1 and Section 3.2] The semantic discrepancy ΔS is computed once from videos generated by the initial SFT policy and is treated as a fixed property of each preference pair throughout training. As the policy changes, the semantic relationship between chosen and rejected videos for the same prompt can change; the weight W(ΔS) may then no longer reflect the constraint for the current policy. Since the theoretical bounds in Propositions 3.1 and Corollary 3.2 treat ΔS as constants, they describe the initial dataset, not the training distribution. The paper should either recompute ΔS periodically, provide an argument for its invariance, or explicitly reframe the analysis as a per-dataset heuristic rather than a guarantee about training.
  5. [Table 6 and Section 3.3] The empirical temperature study shows that with τ=0.10, the VBench Semantics score (73.61) is substantially lower than vanilla DPO (77.01), contradicting the general claim that PSDPO's modulation preserves semantic consistency. The monotonic relationship between τ and semantic preservation is not demonstrated; the reported numbers have a non-monotonic pattern (τ=0.05 gives 76.22, τ=0.02 gives 77.08). This suggests the theoretical statement that τ controls residual drift continuously is not directly reflected in the metric, and the relationship may be confounded by other training dynamics.
minor comments (8)
  1. [Eq. (3) and Eq. (4)] Define ΔS explicitly before first use in Eq. (3); currently ΔS is introduced in prose but not in the equation, and Eq. (4) uses W(ΔS) without a definition nearby.
  2. [Table 3] The ProPhy row reports only Overall; Hard, Activity, and Interaction are missing. The comparison would be more informative if the full subset scores were provided or clearly marked as not available.
  3. [Figure 1 caption] The caption contains stray text fragments 'Incorrect number' and 'Incorrect releasing motion' that appear to be artifacts and should be removed or rewritten.
  4. [Section 3.1] The text says 'approximately 51.50%' while Figure 2 says 'more than half'. Use the same statistic or clarify that these refer to the same quantity.
  5. [Table 6] The caption uses λ_sem (introducing semantic asymmetric modulation) while the text uses τ; unify the notation to avoid confusion between the modulation scale and the weighting temperature.
  6. [Section 4.2] The last sentence says 'reference policy π_ref is fixed as the SFT-initialized Wan-1.3 model' but the model is referred to as Wan-1.3B elsewhere. Fix the abbreviation.
  7. [Appendix D] The heading says 'Wan2.2-5B model' while the text says 'Wan2.2-TI2V-5B'. Use a consistent model name.
  8. [Section 3.2 / Conclusion] The claim that PSDPO 'requires no auxiliary models' may be misleading because PCE and ViCLIP are used to construct the dataset and compute ΔS. Clarify that no auxiliary models are needed during training or inference of the final policy.

Circularity Check

2 steps flagged

Proposition 3.1 and Corollary 3.2 restate the PSDPO loss reweighting by construction; the claimed 'semantic drift bound' is a bound on the conflict-pair gradient term that Eq. (4) was designed to shrink.

specific steps
  1. self definitional [Section 3.3, Eq. (6)-(7), Proposition 3.1]
    "Under PSDPO, each pair in D_conf is scaled by W(ΔS)=exp(ΔS/τ), yielding: E[∇θL_PSDPO]=E_Dmatch[∇θℓ]+ε, where the residual ε is bounded as follows. Proposition 3.1 (Residual gradient bound). Let G=sup∥∇θℓ∥ denote the uniform gradient bound. Then: ∥ε∥ ≤ G·E_Dconf[exp(ΔS/τ)]."

    ε is not an independently defined measure of semantic drift; it is exactly the D_conf contribution to the PSDPO gradient after multiplying by W(ΔS). The bound ∥ε∥≤G·E[W] is the triangle inequality applied to that defined term. Thus Proposition 3.1 restates the method's own reweighting (Eq. 4) rather than bounding any drift of samples from π_θ. The abstract's phrase 'bounds the semantic drift from conflicting pairs' is this definitional bound, so the theoretical 'prediction' reduces to the construction of the loss.

  2. self definitional [Section 3.4, Corollary 3.2 and Eqs. (8)-(10)]
    "Let E_direct = Σ_{t=1}^{T_total} ε^(t) and E_staged = Σ_{t=T_s+1}^{T_total} ε^(t) denote the cumulative residual under direct and staged training, respectively... By the triangle inequality and Proposition 3.1: ∥E_staged∥ ≤ (T_total−T_s)·B, ∥E_direct∥ ≤ T_total·B. Staged training reduces the cumulative residual bound by a factor of T_s/T_total."

    Eq. (10) defines D^(t)=D_match for t<T_s, so ε^(t)=0 in the first stage by construction. Corollary 3.2 simply sums over the steps where D_conf is included; the 'reduction' is exactly the number of excluded steps. It is a restatement of the staged protocol's definition, and because ε^(t) is itself only the reweighted conflict-pair gradient, it does not provably reduce semantic drift of the trained policy. The 'provable' claim is a counting identity.

full rationale

The paper's empirical core is not circular: PSDPO is trained and scored against external benchmarks (VideoPhy-2, PhyGenBench, VBench), and the PCE judge is validated against human agreement, so the reported gains are real evidence. What is circular are the two theoretical 'guarantees.' In Eq. (6), ε is defined as the D_conf gradient after reweighting; Proposition 3.1 then bounds ε by G·E[W], which is the triangle inequality applied to the method's own construction. This is a bound on the reweighted conflict-pair gradient, not on any independent measure of semantic drift of samples from π_θ. Corollary 3.2 inherits this: because Eq. (10) defines the first T_s steps as D_match-only, ε^(t)=0 there, so the cumulative bound differs by exactly T_s·B; 'provably reduces cumulative drift' is a restatement of the protocol's definition. The constrained problem in Eq. (2) is also posed on fixed dataset videos and contains no θ, so the 'constraint' is a labeling of the reweighting rather than a policy constraint. These are definitional reductions of the central theoretical claims, meriting 6; the external experimental comparisons prevent a higher score.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The method rests on several assumptions about the reliability of the semantic scorer (ViCLIP), the physical-preference labeler (PCE), and the stability of ΔS over training. The free parameters τ, T_s, and β are tuned on the evaluation benchmarks rather than derived.

free parameters (4)
  • τ (semantic temperature) = 0.02
    Selected by ablation (Table 6) for best VBench/VideoPhy-2 balance; controls the exponential suppression W(ΔS)=exp(ΔS/τ).
  • T_s (switching step) = 1000 (T_total=3000)
    Swept over {500,1000,2000} (Figure 7); chosen to balance early stability and second-stage physical signal.
  • β (DPO temperature) = 2000
    Set from empirical observations of video generation stability (Section 4.2), not derived.
  • Mflow static threshold = 0.005
    Used by PCE to classify videos as static; chosen by hand without sensitivity analysis.
axioms (5)
  • domain assumption Uniform gradient bound G = sup_θ ||∇ℓ|| is finite.
    Needed for Proposition 3.1; not verified for Wan-1.3B LoRA training.
  • domain assumption ViCLIP scores provide a valid measure of semantic alignment with the prompt.
    Used to define ΔS and partition D_match/D_conf; no human validation of ViCLIP on this dataset is reported.
  • domain assumption PCE physical preference labels are sufficiently reliable (75.54% human agreement) to serve as ground truth.
    The method's supervision depends on PCE; the 24.46% disagreement rate is not analyzed.
  • ad hoc to paper Per-pair ΔS remains valid as the policy is updated.
    ΔS computed once on SFT-generated pairs; the paper does not recompute or verify the constraint during training.
  • ad hoc to paper Reweighting the DPO loss with W(ΔS) enforces the semantic constraint in Eq. 2.
    The paper asserts a correspondence but does not prove that the optimum of the weighted loss satisfies S_sem(y_w,x) ≥ S_sem(y_l,x).

pith-pipeline@v1.3.0-alltime-deepseek · 18175 in / 17278 out tokens · 157346 ms · 2026-08-01T19:27:51.779884+00:00 · methodology

0 comments
read the original abstract

Text-to-video (T2V) generation models have achieved strong visual realism, but improving physical plausibility can come at the cost of semantic consistency with the input text. This tension arises because physical preference is typically determined by comparing dynamics between two videos, without accounting for whether either video faithfully depicts the scene specified by the prompt, making physical-semantic conflict a systematic tendency under this supervision paradigm. We formulate this challenge as a constrained preference optimization problem and propose Physical and Semantic Direct Preference Optimization (PSDPO), which modulates each preference pair's contribution based on the agreement between its physical and semantic signals. A gradient-level analysis shows that PSDPO bounds the semantic drift from conflicting pairs to a controllable residual, and further motivates a staged optimization protocol that provably reduces cumulative drift. The resulting method operates entirely within the standard DPO framework, requiring no auxiliary models or additional loss terms. Experiments show that PSDPO improves physical plausibility by up to $2\times$ over the baseline on VideoPhy-2, while maintaining strong semantic consistency on VBench, achieving a more reliable balance than existing preference-based methods.

Figures

Figures reproduced from arXiv: 2607.16947 by Ping Liu, Shu Zhang, Siwei Meng, Yawei Luo.

Figure 1
Figure 1. Figure 1: Failure mode of preference-based physical alignment and comparison of preference accuracy. (a) Vanilla DPO is trained using [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Match and conflict pairs in the preference dataset and the semantic score distribution. (a) Preference pairs from the dataset under [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the PSDPO framework. (a) Preference pairs are classified as semantically matching ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The architecture of Physical Commonsense Evaluation. PCE evaluates physical plausibility through a hierarchical pipeline that [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Comparison with alignment-based and physical-aware models. Existing methods often suffer from incorrect physical [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Comparison with baseline and SFT models. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Preference accuracy during training under different [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of DPO Margin. The exponential form [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional optimization curves on Wan2.2-TI2V-5B. (a) Supervised fine-tuning loss decreases smoothly and provides a [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Failure Case Analysis. (Top) Under the implicit prompt, all models fail to infer the color change, revealing the [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 9 linked inside Pith

  1. [1]

    Videophy: Evaluating physical commonsense for video generation.ICLR, 2025

    Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bit- ton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation.ICLR, 2025

    2025

  2. [2]

    Videophy-2: A challenging action-centric physical com- monsense evaluation in video generation.arXiv preprint arXiv:2503.06800, 2025

    Hritik Bansal, Clark Peng, Yonatan Bitton, Roman Goldenberg, Aditya Grover, and Kai-Wei Chang. Videophy-2: A challenging action-centric physical com- monsense evaluation in video generation.arXiv preprint arXiv:2503.06800, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    A note on the inception score.arXiv preprint arXiv:1801.01973, 2018

    Shane Barratt and Rishi Sharma. A note on the inception score.arXiv preprint arXiv:1801.01973, 2018

    Pith/arXiv arXiv 2018

  4. [4]

    Curricu- lum learning

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curricu- lum learning. InICML, 2009

    2009

  5. [5]

    Hierarchical fine-grained preference optimization for physically plausible video generation.NeurIPS, 2025

    Harold Haodong Chen, Haojian Huang, Qifeng Chen, Harry Yang, and Ser-Nam Lim. Hierarchical fine-grained preference optimization for physically plausible video generation.NeurIPS, 2025

    2025

  6. [6]

    Deep reinforcement learning from human preferences.NeurIPS, 2017

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences.NeurIPS, 2017

    2017

  7. [7]

    Understanding world or predicting future? a comprehensive survey of world models.ACM Computing Surveys, 2025

    Jingtao Ding, Yunke Zhang, Yu Shang, Yuheng Zhang, Zefang Zong, Jie Feng, Yuan Yuan, Hongyuan Su, Nian Li, Nicholas Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models.ACM Computing Surveys, 2025

    2025

  8. [8]

    Ltx- video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx- video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. InNeurIPS, 2018

    2018

  10. [10]

    En- hancing physical plausibility in video generation by reasoning the implausibility

    Yutong Hao, Chen Chen, Ajmal Saeed Mian, Chang Xu, and Daochang Liu. En- hancing physical plausibility in video generation by reasoning the implausibility. arXiv preprint arXiv:2509.24702, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. InEMNLP, 2021

    2021

  12. [12]

    Lora: Low-rank adaptation of large language models.ICLR, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 2022

    2022

  13. [13]

    Vistadpo: Video hierarchical spatial-temporal direct preference optimization for large video models.ICML, 2025

    Haojian Huang, Haodong Chen, Shengqiong Wu, Meng Luo, Jinlan Fu, Xinya Du, Hanwang Zhang, and Hao Fei. Vistadpo: Video hierarchical spatial-temporal direct preference optimization for large video models.ICML, 2025

    2025

  14. [14]

    Vbench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. InCVPR, 2024

    2024

  15. [15]

    Physmas- ter: Mastering physical representation for video generation via reinforcement learning.arXiv preprint arXiv:2510.13809, 2025

    Sihui Ji, Xi Chen, Xin Tao, Pengfei Wan, and Hengshuang Zhao. Physmas- ter: Mastering physical representation for video generation via reinforcement learning.arXiv preprint arXiv:2510.13809, 2025

    arXiv 2025

  16. [16]

    Huvidpo: Enhancing video generation through direct preference optimization for human- centric alignment.arXiv preprint arXiv:2502.01690, 2025

    Lifan Jiang, Boxi Wu, Jiahui Zhang, Xiaotong Guan, and Shuang Chen. Huvidpo: Enhancing video generation through direct preference optimization for human- centric alignment.arXiv preprint arXiv:2502.01690, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    How far is video generation from world model: A physical law perspective.ICML, 2025

    Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective.ICML, 2025

    2025

  18. [18]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InICCV, 2021

    2021

  19. [19]

    Hunyuanvideo: A systematic frame- work for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic frame- work for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    What about gravity in video generation? post-training newton’s laws with verifiable rewards.arXiv preprint arXiv:2512.00425, 2025

    Minh-Quan Le, Yuanzhi Zhu, Vicky Kalogeiton, and Dimitris Samaras. What about gravity in video generation? post-training newton’s laws with verifiable rewards.arXiv preprint arXiv:2512.00425, 2025

    arXiv 2025

  21. [21]

    Videodpo: Omni-preference alignment for video diffusion generation

    Runtao Liu, Haoyu Wu, Ziqiang Zheng, Chen Wei, Yingqing He, Renjie Pi, and Qifeng Chen. Videodpo: Omni-preference alignment for video diffusion generation. InCVPR, 2025

    2025

  22. [22]

    Physgen: Rigid-body physics-grounded image-to-video generation

    Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. InECCV, 2024

    2024

  23. [23]

    Evalcrafter: Benchmarking and evaluating large video generation models

    Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. InCVPR, 2024

    2024

  24. [24]

    Decoupled weight decay regularization.ICLR, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.ICLR, 2019

    2019

  25. [25]

    Contextanyone: Context-aware diffusion for character-consistent text-to-video generation.arXiv preprint arXiv:2512.07328, 2025

    Ziyang Mai and Yu-Wing Tai. Contextanyone: Context-aware diffusion for character-consistent text-to-video generation.arXiv preprint arXiv:2512.07328, 2025

    arXiv 2025

  26. [26]

    Towards world simulator: Crafting physical commonsense-based benchmark for video generation.ICML, 2025

    Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation.ICML, 2025

    2025

  27. [27]

    Grounding creativity in physics: A brief survey of physical priors in aigc.IJCAI, 2025

    Siwei Meng, Yawei Luo, and Ping Liu. Grounding creativity in physics: A brief survey of physical priors in aigc.IJCAI, 2025

    2025

  28. [28]

    Phymagic: Physical motion-aware gener- ative inference with confidence-guided llm.ECCV, 2026

    Siwei Meng, Yawei Luo, and Ping Liu. Phymagic: Physical motion-aware gener- ative inference with confidence-guided llm.ECCV, 2026

    2026

  29. [29]

    Video generation models as world simulators

    OpenAI. Video generation models as world simulators. https://openai .com/ research/video-generation-models-as-world-simulators, 2024. Accessed 2024

    2024

  30. [30]

    Rdpo: Real data preference optimization for physics consistency video generation

    Wenxu Qian, Chaoyue Wang, Hou Peng, Zhiyu Tan, Hao Li, and Anxiang Zeng. Rdpo: Real data preference optimization for physics consistency video generation. arXiv preprint arXiv:2506.18655, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021

    2021

  32. [32]

    Direct preference optimization: Your language model is secretly a reward model.NeurIPS, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.NeurIPS, 2023

    2023

  33. [33]

    Learning to reweight examples for robust deep learning

    Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. InICML, 2018

    2018

  34. [34]

    Physmotion: Physics-grounded dynamics from a single image.3DV, 2026

    Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chen- fanfu Jiang. Physmotion: Physics-grounded dynamics from a single image.3DV, 2026

    2026

  35. [35]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InECCV, 2020

    2020

  36. [36]

    Fvd: A new metric for video generation

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. ICLRW, 2019

    2019

  37. [37]

    Wan: Open and advanced large- scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large- scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Wisa: World simulator assistant for physics-aware text-to-video generation.NeurIPS, 2025

    Jing Wang, Ao Ma, Ke Cao, Jun Zheng, Zhanjie Zhang, Jiasong Feng, Shanyuan Liu, Yuhang Ma, Bo Cheng, Dawei Leng, et al. Wisa: World simulator assistant for physics-aware text-to-video generation.NeurIPS, 2025

    2025

  39. [39]

    Physcorr: Dual-reward dpo for physics- constrained text-to-video generation with automated preference selection.arXiv preprint arXiv:2511.03997, 2025

    Peiyao Wang, Weining Wang, and Qi Li. Physcorr: Dual-reward dpo for physics- constrained text-to-video generation with automated preference selection.arXiv preprint arXiv:2511.03997, 2025

    arXiv 2025

  40. [40]

    A survey on curriculum learning

    Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  41. [41]

    Internvid: A large-scale video-text dataset for multimodal understanding and generation.ICLR, 2024

    Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation.ICLR, 2024

    2024

  42. [42]

    Densedpo: Fine-grained temporal preference optimization for video diffusion models.NeurIPS, 2025

    Ziyi Wu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Ashkan Mirzaei, Igor Gilitschenski, Sergey Tulyakov, and Aliaksandr Siarohin. Densedpo: Fine-grained temporal preference optimization for video diffusion models.NeurIPS, 2025

    2025

  43. [43]

    Physgaussian: Physics-integrated 3d gaussians for generative dynamics

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. InCVPR, 2024

    2024

  44. [44]

    VideoCLIP: Contrastive pre-training for zero-shot video-text understanding.EMNLP, 2021

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pre-training for zero-shot video-text understanding.EMNLP, 2021

    2021

  45. [45]

    Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation

    Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation. InCVPR, 2025

    2025

  46. [46]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Heng- shuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InCVPR, 2024

    2024

  48. [48]

    Cogvideox: Text-to-video diffusion models with an expert transformer.ICLR, 2025

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.ICLR, 2025

    2025

  49. [49]

    Inference-time physics alignment of video gener- ative models with latent world models.arXiv preprint arXiv:2601.10553, 2026

    Jianhao Yuan, Xiaofeng Zhang, Felix Friedrich, Nicolas Beltran-Velez, Melissa Hall, Reyhane Askari-Hemmat, Xiaochuang Han, Nicolas Ballas, Michal Drozdzal, and Adriana Romero-Soriano. Inference-time physics alignment of video gener- ative models with latent world models.arXiv preprint arXiv:2601.10553, 2026

    arXiv 2026

  50. [50]

    Think before you diffuse: Llms-guided physics-aware video generation.arXiv preprint arXiv:2505.21653, 2025

    Ke Zhang, Cihan Xiao, Yiqun Mei, Jiacong Xu, and Vishal M Patel. Think before you diffuse: Llms-guided physics-aware video generation.arXiv preprint arXiv:2505.21653, 2025

    arXiv 2025

  51. [51]

    Physrvg: Physics- aware unified reinforcement learning for video generative models.arXiv preprint arXiv:2601.11087, 2026

    Qiyuan Zhang, Biao Gong, Shuai Tan, Zheng Zhang, Yujun Shen, Xing Zhu, Yuyuan Li, Kelu Yao, Chunhua Shen, and Changqing Zou. Physrvg: Physics- aware unified reinforcement learning for video generative models.arXiv preprint arXiv:2601.11087, 2026

    arXiv 2026

  52. [52]

    Judging llm-as-a-judge with mt-bench and chatbot arena.NeurIPS, 2023

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.NeurIPS, 2023

    2023