Pith. sign in

REVIEW 3 major objections 4 minor 74 references

Physics-Grounded Motion Forecasting via Equation Discovery for Trajectory-Guided Image-to-Video Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a symbolic-regression method called ReSR can recover the equations of motion from short video clips, then use those equations to forecast trajectories that make image-to-video generation track physically plausible…

desk verdict Useful pipeline plumbing, but the equation-discovery headline does not survive contact with the paper's own equation bank. read the letter →

arxiv 2507.06830 v1 pith:3EEMWDRJ submitted 2025-07-09 cs.CV cs.AI

Tao Feng , Xianbing Zhao , Zhenhua Chen , Tien Tsin Wong , Hamid Rezatofighi , Gholamreza Haffari , Lizhen Qu This is my paper
classification cs.CVcs.AI
keywords physics-groundedvideogenerationsymbolicregressionequationdiscoveryretrieval-basedpre-trainingtrajectory-guidedimage-to-videomotionforecastingclassicalmechanicsunderstanding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a video's object motion can be turned back into an explicit equation of motion by symbolic regression, and that the equation improves video generation when used as a trajectory guide. The proposed method, ReSR, seeds the symbolic search with equations retrieved from a bank of physics formulas, selected by shape similarity to the observed trajectory. In classical-mechanics test videos such as spring-mass, pendulum, and projectile motion, the paper reports that ReSR recovers equations close to the known analytical forms and that trajectory-guided video models produce future frames that follow the predicted paths more closely than baselines. The method is inference-only: existing image-to-video models are used as-is, without fine-tuning.

What carries the argument

The load-bearing mechanism is ReSR's retrieval-based initialization inside an evolutionary symbolic regression search. Given a trajectory extracted from the video, ReSR normalizes the trajectory and trajectories sampled from each equation in a physics equation bank, compares them with normalized dynamic time warping, and seeds a fraction $\alpha$ of the initial evolutionary population with the top-$k$ most shape-similar equations; the remainder is random. This turns the equation bank into a prior for the genetic search, accelerating convergence. The discovered equation pair $x_t=f^x_i(t)$, $y_t=f^y_i(t)$ then extrapolates positions beyond the observed frames and becomes the conditioning trajectory for an image-to-video model.

What would settle it

Delete the manually added damped-harmonic-oscillator and projectile equations from the bank, rerun ReSR on the same spring-mass and projectile videos, and check whether the discovered equations still match the ground truth on tree-edit distance and trajectory error; if performance collapses to baseline levels, the recovery result depends on bank membership rather than on the search itself.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that retrieval-based pre-training makes evolutionary symbolic regression fast and accurate enough to discover motion equations from a handful of video trajectories, and that those equations, rather than hand-drawn or statistically estimated paths, should supply the control signal for trajectory-guided image-to-video models. The paper finds that ReSR attains lower tree-edit distance to ground-truth equations and lower trajectory error than the compared symbolic-regression baselines, and that a trajectory-guided model steered by ReSR's forecast paths yields videos judged more physically aligned than the same model steered by manually drawn trajectories or by a strong symbolic-regression baseline's trajectories. The paper also reports that even the best trajectory-guided models remain far behind a rule-based physics simulator on trajectory error, so the approach narrows but does not close the gap between generative video and physical simulation.

Load-bearing premise

The evaluation assumes that every target system's correct equation is present in the equation bank; if the damped-oscillator and projectile formulas were removed from the bank, the claimed recovery could fail.

Editorial extensions

If this is right

  • If ReSR is right, video generation can be made physics-aware at inference time for any trajectory-guided image-to-video model, with no retraining or fine-tuning.
  • Discovered equations generalize beyond the observed frames, so forecast trajectories of arbitrary length follow the same discovered law rather than a fitted curve.
  • Equation recovery outputs human-interpretable symbolic expressions, unlike latent-dynamics or learned-trajectory alternatives.
  • Even with physics-aligned trajectories, current image-to-video models remain far from simulator-level fidelity, so trajectory guidance is necessary but not sufficient for physical consistency.
  • The same pipeline could be extended to other motion types whenever object tracks can be extracted and the equation bank contains a representative family.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that removing the manually added oscillator and projectile equations from the bank and rerunning the same evaluation would separate genuine discovery from retrieval of a known candidate.
  • If equation-bank coverage is the active ingredient, the method's practical ceiling is bounded by bank completeness; a natural next step is to make the bank self-extending by adding newly discovered equations back into it.
  • The normalized dynamic time warping idea could carry over to other sequence-conditioning tasks where only shape, not scale or offset, matters, such as gesture or action trajectory transfer.
  • For real-world deployment, the orthographic-view and full-observability assumptions would need relaxing, since projective camera effects and occlusions would introduce noise the current pipeline does not model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a framework for physics-grounded video generation that combines symbolic regression (SR) with trajectory-guided image-to-video (I2V) models. It introduces ReSR, a retrieval-based pre-training mechanism for SR, which initializes the evolutionary population with equations retrieved from a curated equation bank using a normalized dynamic time warping (N-DTW) similarity measure. The discovered equations are used to predict future trajectories, which condition an off-the-shelf I2V model. The paper reports that ReSR recovers ground-truth analytical equations on classical mechanics benchmarks and that the resulting trajectories improve the physical alignment of generated videos over baselines.

Significance. If substantiated, the paper would make a useful contribution by coupling interpretable equation discovery with video generation, and the retrieval-based pre-training idea is a practical way to accelerate SR. The authors provide implementation details via a modified SymbolicRegression.jl framework, enhancing reproducibility, and the evaluation covers a broad set of I2V baselines including commercial models. However, the central discovery claim is not yet firmly established because the equation bank contains the target equations for the evaluated systems; the improvements on systems without analytical solutions (single/double pendulum, fluid) are not subject to this contamination, which partially mitigates the concern for the video-generation results, but the headline claim of recovering ground-truth equations is still based on the contaminated subset.

major comments (3)
  1. [3.3 and 4.1 (Table 2)] The equation bank described in Section 3.3 is augmented with 'manually augmented physics equations' covering 'damped harmonic oscillators and projectile motions,' which are exactly the ground-truth systems evaluated in Section 4.1 (spring-mass, damped spring-mass, projectile). Since ReSR seeds 75% of its initial population (alpha=0.75) with top-k retrieved equations and N-DTW will retrieve the matching functional forms, the high TED values in Table 2 likely reflect retrieval plus parameter fitting rather than de novo discovery. The ablation showing ReSR-0 identical to PySR and monotonic improvement with alpha is consistent with this interpretation. To support the discovery claim, the authors should either remove the target equations from the bank or report the proportion of final equations that are direct descendants of retrieved seeds, and demonstrate that ReSR can still discover the ground-truth equations from primitive operators alone.
  2. [4.2 (Tables 3 and 4)] The automatic video metrics in Table 3 are presented as averages without error bars or significance tests, and the human evaluation in Table 4 reports win rates from only three annotators with no inter-annotator agreement or confidence intervals. Under these conditions, small differences (e.g., TrajErr 325 vs 357 for synthetic Kling-ReSR vs Kling-Manual) are not interpretable. Please provide bootstrap confidence intervals or paired significance tests for the automatic metrics, and report per-annotator agreement (e.g., Fleiss' kappa) or at least individual counts for the pairwise preferences.
  3. [3.3] The retrieval-based pre-training has two free parameters that are not controlled: the number of retrieved equations k (never specified) and the constant-substitution strategy of replacing non-time variables with 10 in the Feynman and Nguyen equations. Without reporting k and without ablating the constant-substitution choice, it is unclear whether the gains of ReSR are sensitive to bank construction. The authors should report the value of k and perform a sensitivity analysis over both k and the substitution constant.
minor comments (4)
  1. [2] The text contains a typo: 'obercational data' should be 'observational data'.
  2. [1 (author affiliation)] The affiliation for Xianbing Zhao lists 'Shezhen' as the city; this should be 'Shenzhen'.
  3. [4.2 (Table 3)] The caption states that 'all reported metrics are averaged across physical systems,' but the table separately reports real and synthetic initial frames; please clarify whether the values are averages within each setting or global averages across all systems.
  4. [Related Work] Reference [25] (Huang et al.) also learns continuous dynamics from videos and is used as the data source; the related work section should explicitly discuss the differences between that approach and the proposed ReSR, particularly the focus on symbolic equations and retrieval-based initialization.

Circularity Check

1 steps flagged · score 6.0 of 10

ReSR's reported 'recovery of ground-truth equations' is partly retrieval from an equation bank that was manually built to contain those same equations.

  1. fitted input called prediction [§3.3 (Construction of Equation Bank; Initialization of ReSR), §4.1 (Datasets; Table 2)]
    "A set of manually augmented physics equations not included in the above datasets. These cover well-known motion dynamics such as (damped) harmonic oscillators and projectile motions, ensuring the equation bank includes representative equations for various physical systems. ... systems with ground-truth trajectory equations (i.e., systems with analytical solutions), including spring mass, damped spring mass, two body, and projectile motion."

    The §4.1 targets—spring-mass, damped spring-mass, and projectile motion—are the same functional families that §3.3 manually adds to the bank. ReSR selects top-k N-DTW equations and seeds the initial population with them; at the best setting (α=0.75), 75% of the population begins from retrieved equations. The TED scores in Table 2 therefore measure parameter fitting of an already-present functional form, not discovery from primitive operators. The experiment cannot distinguish retrieval of a pre-inserted answer from discovery, so the headline recovery claim is partly forced by construction.

full rationale

The circularity is limited to the symbolic-regression evidence backing the 'discovery' claim. The video-generation evaluation (Tables 3 and 4) is not circular: it compares trajectory-guided I2V models on external metrics and human preference, and those results are independent of whether the equation bank contains the target systems. However, because the bank's manually augmented equations coincide with the ground-truth systems in §4.1 and because ReSR's initialization is dominated by retrieved candidates, Table 2 cannot support the strong claim that ReSR discovers ground-truth equations from scratch. A clean test would remove the target equations from the bank or compare against a bank without them; the current design leaves the recovery result partly predetermined. No load-bearing self-citation or uniqueness-importation issue was found.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a trajectory extraction heuristic and on an equation bank curated to include the ground-truth solutions of the test systems. The alpha hyperparameter and the constant used to replace time-independent physical variables are chosen by hand. No new physical entities are introduced.

free parameters (3)
  • init_weight_alpha = 0.75
    Controls the fraction of the initial symbolic regression population seeded with retrieved bank equations; tuned on validation and reported as optimal in Table 2.
  • top_k_trajectories = 5
    Number of highest-variance trajectories chosen as inputs for equation learning; a heuristic that assumes the object dominates motion.
  • equation_bank_constant_substitution = 10
    Time-independent physical variables (mass, density) are replaced with the constant 10 when adapting Feynman/Nguyen equations, an arbitrary choice that affects retrieved candidates.
assumptions (4)
  • domain assumption Object motion is confined to a 2D plane, recorded from an orthographic viewpoint, and trajectories are fully observable across the video.
    Stated in Section 4, this restricts the method to simple scenes and limits generalization to perspective, occluded, or 3D motion.
  • domain assumption CoTracker, with uniform grid query points and top-variance selection, yields trajectories that accurately isolate the moving object.
    Section 3.2 assumes the most dynamic points are the object, which may fail for multiple moving objects, camera motion, or deformable objects.
  • ad hoc to paper The equation bank contains equations sufficient to cover the test systems, including the ground-truth analytical solutions.
    Section 3.3 adds manual equations for damped harmonic oscillators and projectiles, matching the evaluation systems; this is the main source of circularity in the equation recovery claim.
  • domain assumption N-DTW shape similarity is a valid proxy for matching a trajectory to its governing equation family.
    Section 3.3 relies on normalized DTW to retrieve candidates; no analysis is provided on when this retrieval fails or on sensitivity to noise and phase.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Grounded Motion Forecasting via Equation Discovery for Trajectory-Guided Image-to-Video Generation." pith.science (2026). https://pith.science/paper/3EEMWDRJ

@misc{pith2026250706830,
  author       = {Pith},
  title        = {Pith review of: Physics-Grounded Motion Forecasting via Equation Discovery for Trajectory-Guided Image-to-Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3EEMWDRJ}},
  note         = {Machine review of arXiv:2507.06830}
}
read the original abstract

Recent advances in diffusion-based and autoregressive video generation models have achieved remarkable visual realism. However, these models typically lack accurate physical alignment, failing to replicate real-world dynamics in object motion. This limitation arises primarily from their reliance on learned statistical correlations rather than capturing mechanisms adhering to physical laws. To address this issue, we introduce a novel framework that integrates symbolic regression (SR) and trajectory-guided image-to-video (I2V) models for physics-grounded video forecasting. Our approach extracts motion trajectories from input videos, uses a retrieval-based pre-training mechanism to enhance symbolic regression, and discovers equations of motion to forecast physically accurate future trajectories. These trajectories then guide video generation without requiring fine-tuning of existing models. Evaluated on scenarios in Classical Mechanics, including spring-mass, pendulums, and projectile motions, our method successfully recovers ground-truth analytical equations and improves the physical alignment of generated videos over baseline methods.

Figures

Figures reproduced from arXiv: 2507.06830 by the authors.

Figure 1
Figure 1. An overview of our proposed framework. Given an input video, we first extract object ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Convergence comparison between our method and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparison of observed trajectories with ground [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons across models, showcasing [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 44 canonical work pages

  1. [1]

    Jeff Bezanson, Alan Edelman, Stefan Karpinski, and Viral B Shah. 2017. Julia: A fresh approach to numerical computing.SIAM review59, 1 (2017), 65–98. https://doi.org/10.1137/141000671

    doi:10.1137/141000671 2017

  2. [2]

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Giambat- tista Parascandolo. 2021. Neural symbolic regression that scales. InInternational Conference on Machine Learning. Pmlr, 936–945

    work page 2021

  3. [3]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. 2023. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127(2023)

    arXiv 2023

  4. [4]

    Florent Bonnet, Jocelyn Ahmed Mazari, Paola Cinnella, and patrick gallinari. 2022. AirfRANS: High Fidelity Computational Fluid Dynamics Dataset for Approxi- mating Reynolds-Averaged Navier–Stokes Solutions. InThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    work page 2022

  5. [5]

    Anne Brindle. 1980. Genetic algorithms for function optimization. (1980)

    work page 1980

  6. [6]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging Properties in Self-Supervised Vision Transformers. In2021 IEEE/CVF International Conference on Computer Vision (ICCV). 9630–9640. doi:10.1109/ICCV48922.2021.00951

    arXiv 2021

  7. [7]

    Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. 2023. StableVideo: Text- driven Consistency-aware Diffusion Video Editing . In2023 IEEE/CVF Inter- national Conference on Computer Vision (ICCV). IEEE Computer Society, Los Alamitos, CA, USA, 22983–22993. doi:10.1109/ICCV51070.2023.02106

    arXiv 2023

  8. [8]

    Pradyumna Chari, Chinmay Talegaonkar, Yunhao Ba, and Achuta Kadambi. 2019. Visual Physics: Discovering Physical Laws from Videos. arXiv:1911.11893 [cs.CV] https://arxiv.org/abs/1911.11893

    arXiv 2019

Show all 74 references
  1. [9]

    Boyuan Chen, Kuang Huang, Sunand Raghupathi, Ishaan Chandratreya, Qiang Du, and Hod Lipson. 2022. Automated discovery of fundamental variables hidden in experimental data.Nature Computational Science2, 7 (2022), 433–442

    2022

  2. [11]

    Miles Cranmer. 2023. Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl. arXiv:2305.01582 [astro-ph.IM] https://arxiv.org/abs/ 2305.01582

    2023 arXiv

  3. [12]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human...

    2019

  4. [13]

    Ilias Dimoulkas, Lars Herre, Dina Khastieva, Elis Nycander, Mikael Amelin, and Peyman Mazidi. 2018. A Hybrid Model Based on Symbolic Regression and Neural Networks for Electricity Load Forecasting. In2018 15th International Conference on the European Energy Market (EEM). 1–5. ...

    2018

  5. [14]

    Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. 2010. Why Does Unsupervised Pre-training Help Deep Learning?. InProceedings of the Thirteenth International Conference on Artificial Intelligence and Statis- tics (Proceedings of Machine Learning Research, Vol....

    2010

  6. [15]

    Tao Feng, Lizhen Qu, and Gholamreza Haffari. 2023. Less is More: Mitigate Spurious Correlations for Open-Domain Dialogue Response Generation Models by Causal Discovery.Transactions of the Association for Computational Linguistics 11 (2023), 511–530. doi:10.1162/tacl_a_00561

    2023 doi

  7. [16]

    Tao Feng, Lizhen Qu, Xiaoxi Kang, and Gholamreza Haffari. 2025. CausalScore: An Automatic Reference-Free Metric for Assessing Response Relevance in Open- Domain Dialogue Systems. InProceedings of the 31st International Conference on Computational Linguistics, Owen Rambow, Leo ...

    2025

  8. [17]

    Tao Feng, Lizhen Qu, Zhuang Li, Haolan Zhan, Yuncheng Hua, and Reza Haf

  9. [18]

    Feynman, R.B

    R.P. Feynman, R.B. Leighton, M. Sands, and EM Hafner. 1965.The Feynman Lectures on Physics. AAPT. 750 pages

    1965

  10. [19]

    Xiao FU, Xian Liu, Xintao Wang, Sida Peng, Menghan Xia, Xiaoyu Shi, Ziyang Yuan, Pengfei Wan, Di ZHANG, and Dahua Lin. 2025. 3DTrajMaster: Mastering 3D Trajectory for Multi-Entity Motion in Video Generation. InThe Thirteenth International Conference on Learning Representations...

    2025

  11. [20]

    Alejandro Castañeda Garcia, Jan van Gemert, Daan Brinks, and Nergis Tömen

  12. [21]

    David E Goldberg and Kalyanmoy Deb. 1991. A comparative analysis of selection schemes used in genetic algorithms. InFoundations of genetic algorithms. Vol. 1. Elsevier, 69–93

    1991

  13. [22]

    arXiv:2410.01376 [cs.CV] https://arxiv.org/ abs/2410.01376

    Learning Physics From Video: Unsupervised Physical Parameter Estimation for Continuous Dynamical Systems. arXiv:2410.01376 [cs.CV] https://arxiv.org/ abs/2410.01376

    arXiv

  14. [23]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. InAdvances in Neural Information Processing Systems, I. Guyon, U. Von Luxburg, S. Bengio, H. Wall...

    2017

  15. [24]

    Arya Grayeli, Atharva Sehgal, Omar Costilla Reyes, Miles Cranmer, and Swarat Chaudhuri. 2024. Symbolic Regression with a Learned Concept Library. InICML 2024 AI for Science Workshop. https://openreview.net/forum?id=2vZ411eb1j

    2024

  16. [25]

    Kuang Huang, Dong Heon Cho, and Boyuan Chen. 2024. Automated Discovery of Continuous Dynamics from Videos. arXiv:2410.11894 [eess.SY] https://arxiv. org/abs/2410.11894

    2024 arXiv

  17. [26]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. InProceedings of the 34th International Conference on Neural Information Processing Systems(Vancouver, BC, Canada)(NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, Article 574, 12 pages

    2020

  18. [27]

    Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. 2024. How Far is Video Generation from World Model: A Physical Law Perspective. arXiv:2411.02385 [cs.CV] https://arxiv.org/abs/2411.02385

    2024 arXiv

  19. [28]

    Pierre-Alexandre Kamienny, Stéphane d’Ascoli, Guillaume Lample, and Fran- cois Charton. 2022. End-to-end Symbolic Regression with Transformers. In Advances in Neural Information Processing Systems, Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (Eds.). https:...

    2022

  20. [29]

    Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. 2024. CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos. arXiv:2410.11831 [cs.CV] https://arxiv.org/ abs/2410.11831

    2024 arXiv

  21. [30]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models. arXiv:2001.08361 [cs.LG] https: //arxiv.org/abs/2001.08361

    2020 arXiv

  22. [31]

    Georg Kohl, Liwei Chen, and Nils Thuerey. 2024. Turbulent Flow Simulation using Autoregressive Conditional Diffusion Models

    2024

  23. [32]

    Byung Chun Kim, Byungro Kim, and Yoonsuk Hyun. 2024. Investigation of out-of-distribution detection across various models and training methodologies. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al. Neural Networks175 (2024), 106288

    2024

  24. [33]

    Santiago, Ignacio Aravena, Terrell N

    Mikel Landajuela, Chak Lee, Jiachen Yang, Ruben Glatt, Claudio P. Santiago, Ignacio Aravena, Terrell N. Mundhenk, Garrett Mulcahy, and Brenden K. Petersen

  25. [34]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, J...

    2025 arXiv

  26. [35]

    Hou, and Max Tegmark

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljacic, Thomas Y. Hou, and Max Tegmark. 2025. KAN: Kolmogorov–Arnold Networks. InThe Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=Ozo7qJ5vZi

    2025

  27. [36]

    Lele Luan, Yang Liu, and Hao Sun. 2021. Uncovering Closed-form Governing Equations of Nonlinear Dynamics from Videos. arXiv:2106.04776 [cs.LG]

    2021 arXiv

  28. [37]

    Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming- Ming Cheng. 2023. Amt: All-pairs multi-field transforms for efficient frame interpolation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9801–9810

    2023

  29. [38]

    Reddy, and Amir Barati Farimani

    Kazem Meidani, Parshin Shojaee, Chandan K. Reddy, and Amir Barati Farimani

  30. [39]

    2007.Game physics engine development

    Ian Millington. 2007.Game physics engine development. CRC Press

    2007

  31. [40]

    Kazem Meidani and Amir Barati Farimani. 2023. Identification of parametric dynamical systems using integer programming.Expert Systems with Applications 219 (2023), 119622

    2023

  32. [41]

    Meinard Müller. 2007. Dynamic time warping.Information retrieval for music and motion(2007), 69–84

    2007

  33. [42]

    InThe Twelfth International Conference on Learning Representations

    SNIP: Bridging Mathematical Symbolic and Numeric Realms with Unified Pre-training. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=KZSEgJGPxu

  34. [43]

    Dalziel, Drummond Buschman Fielding, Daniel Fortunato, Jared A

    Ruben Ohana, Michael McCabe, Lucas Thibaut Meyer, Rudy Morel, Fruzsina Julia Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B. Dalziel, Drummond Buschman Fielding, Daniel Fortunato, Jared A. Goldberg, Keiya Hi- rashima, Yan-Fei Jiang, Rich Kerswell, Suryanar...

    2024

  35. [44]

    Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. 2025. Do generative video models understand physical principles? arXiv:2501.09038 [cs.CV] https://arxiv.org/abs/2501.09038

    2025 arXiv

  36. [45]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. https: //api.semanticscholar.org/CorpusID:160025533

    2019

  37. [46]

    Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B. Lindell. 2025. SG-I2V: Self-Guided Trajectory Control in Image- to-Video Generation. InThe Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=uQjySppU9x

    2025

  38. [47]

    Rudy, Steven L

    Samuel H. Rudy, Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. 2017. Data-driven discovery of partial differential equations. Science Advances3, 4 (2017), e1602614. doi:10.1126/sciadv.1602614 arXiv:https://www.science.org/doi/pdf/10.1126/sciadv.1602614

    2017 doi

  39. [48]

    M. Planck. 1900.Ueber eine Verbesserung der Wien’schen Spectral-Gleichung. J.A. Barth. https://books.google.com.au/books?id=v3ptnQEACAAJ

    1900

  40. [49]

    Xiaoqian Shen and Mohamed Elhoseiny. 2023. StoryGPT-V: Large Language Models as Consistent Story Visualizers. arXiv:2312.02252 [cs.CV] https://arxiv. org/abs/2312.02252

    2023 arXiv

  41. [50]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10674–10685. doi:10.1109/CVPR52688.2022.01042

    2022

  42. [51]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456(2020)

    2020 arXiv

  43. [52]

    Schmidt and Hod Lipson

    Michael D. Schmidt and Hod Lipson. 2010. Age-fitness pareto optimization. In Proceedings of the 12th Annual Conference on Genetic and Evolutionary Computa- tion(Portland, Oregon, USA)(GECCO ’10). Association for Computing Machinery, New York, NY, USA, 543–544. doi:10.1145/1830...

    2010

  44. [53]

    Kling Team. [n. d.]. Kling AI. https://app.klingai.com/cn/en. Accessed: 2023-10- 05

    2023

  45. [54]

    Parshin Shojaee, Kazem Meidani, Amir Barati Farimani, and Chandan K. Reddy

  46. [55]

    Erina Tetriyani, Asep Jihad, Tika Karlina Rachmawati, and Hamdan Sugilar. 2024. Development of Video Animation Media for Learning a System Two-Variable Linear Equation.KnE Social Sciences9, 8 (Apr. 2024), 423–430. doi:10.18502/kss. v9i8.15575

    2024 doi

  47. [56]

    Emanuel Todorov, Tom Erez, and Yuval Tassa. 2012. MuJoCo: A physics engine for model-based control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. 5026–5033. doi:10.1109/IROS.2012.6386109

    2012

  48. [57]

    Trevor Stephens. 2025. gplearn: Genetic Programming in Python. https://github. com/trevorstephens/gplearn Accessed: 18-Mar-2025

    2025

  49. [58]

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. 2019. Towards Accurate Generative Models of Video: A New Metric & Challenges. arXiv:1812.01717 [cs.CV] https://arxiv. org/abs/1812.01717

    2019 arXiv

  50. [59]

    NVIDIA Team. 2025. Cosmos World Foundation Model Platform for Physical AI. arXiv:2501.03575 [cs.CV]

    2025 arXiv

  51. [60]

    Marco Virgolin and Solon P Pissis. 2022. Symbolic Regression is NP-hard.Trans- actions on Machine Learning Research(2022). https://openreview.net/forum?id= LTiaPxqe2e

    2022

  52. [61]

    Jing Wang, Ao Ma, Ke Cao, Jun Zheng, Zhanjie Zhang, Jiasong Feng, Shanyuan Liu, Yuhang Ma, Bo Cheng, Dawei Leng, Yuhui Yin, and Xiaodan Liang. 2025. WISA: World Simulator Assistant for Physics-Aware Text-to-Video Generation. arXiv:2503.08153 [cs.CV] https://arxiv.org/abs/2503.08153

    2025 arXiv

  53. [62]

    Silviu-Marian Udrescu and Max Tegmark. 2020. AI Feynman: A physics-inspired method for symbolic regression.Science advances6, 16 (2020), eaay2631

    2020

  54. [63]

    Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. 2024. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. InACM SIGGRAPH 2024 Conference Papers (Denver, CO, USA)(SIGGRAPH ’24). Association for Compu...

    2024

  55. [64]

    Nguyen Quang Uy, Nguyen Xuan Hoai, Michael O’Neill, Robert I McKay, and Edgar Galván-López. 2011. Semantically-based crossover in genetic program- ming: application to real-valued symbolic regression.Genetic Programming and Evolvable Machines12 (2011), 91–119

    2011

  56. [65]

    Jinbo Xing, Long Mai, Cusuh Ham, Jiahui Huang, Aniruddha Mahapatra, Chi- Wing Fu, Tien-Tsin Wong, and Feng Liu. 2025. MotionCanvas: Cinematic Shot Design with Controllable Image-to-Video Generation. arXiv:2502.04299 [cs.CV] https://arxiv.org/abs/2502.04299

    2025 arXiv

  57. [66]

    Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. 2021. VideoGPT: Video Generation using VQ-VAE and Transformers. arXiv:2104.10157 [cs.CV] https://arxiv.org/abs/2104.10157

    2021 arXiv

  58. [67]

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jin- sheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yu- long Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin,...

    2024 arXiv

  59. [68]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Yux- uan.Zhang, Weihan Wang, Yean Cheng, Bin Xu, Xiaotao Gu, Yuxiao Dong, and Jie Tang. 2025. CogVideoX: Text-to-Video Diffusion Models wi...

    2025

  60. [69]

    Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. 2025. DragAnything: Motion Control for Anything Using Entity Representation. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, O...

    2025

  61. [70]

    Zhenghao Zhang, Junchao Liao, Menghao Li, Zuozhuo Dai, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. 2024. Tora: Trajectory-oriented Diffusion Transformer for Video Generation. arXiv:2407.21705 [cs.CV] https://arxiv.org/ abs/2407.21705

    2024 arXiv

  62. [72]

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. 2024. Generalized out- of-distribution detection: A survey.International Journal of Computer Vision132, 12 (2024), 5635–5662

    2024

  63. [74]

    Kaizhong Zhang and Dennis Shasha. 1989. Simple Fast Algorithms for the Editing Distance between Trees and Related Problems.SIAM J. Comput.18, 6 (1989), 1245–1262. doi:10.1137/0218082 arXiv:https://doi.org/10.1137/0218082

    1989 doi

  64. [2022]

    InAdvances in Neural Information Processing Systems, Alice H

    A Unified Framework for Deep Symbolic Regression. InAdvances in Neural Information Processing Systems, Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (Eds.). https://openreview.net/forum?id=2FNnBhwJsHK

  65. [2023]

    InThirty-seventh Conference on Neural Information Processing Systems

    Transformer-based Planning for Symbolic Regression. InThirty-seventh Conference on Neural Information Processing Systems. https://openreview.net/ forum?id=0rVXQEeFEL

  66. [2024]

    InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.)

    IMO: Greedy Layer-Wise Sparse Representation Learning for Out-of- Distribution Text Classification with Pre-trained Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek S...

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.