pith. sign in

arxiv: 2606.20707 · v1 · pith:7WUY4FTYnew · submitted 2026-06-15 · 💻 cs.CV · cs.AI

GEOPHYS: The Geometry of Physical Plausibility

Christian Intern\`o , Alexander Pondaven , Habon Issa , Fabio Pizzati , Francesco Pinto , Markus Olhofer , Ivan Laptev , Philip Torr
show 3 more authors
Eero P. Simoncelli Barbara Hammer David Klindt
This is my paper

Pith reviewed 2026-06-27 03:32 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords physical plausibilityvideo analysisimage encodersgeometric propertiesphysics violation detectionobject permanencevideo generationfrozen encoders
0
0 comments X

The pith

Five geometric properties of embeddings from frozen image encoders detect physically implausible videos.

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

The paper establishes that physical plausibility indicators emerge as five geometric properties in the per-frame embeddings of ordinary frozen image encoders. These properties correlate with human EEG responses to object-permanence violations. Aggregated into GEOPHYS, the signals separate implausible videos from realistic ones at 98.3 percent accuracy on LikePhys and 93.3 percent on IntPhys2, while large multimodal models and video diffusion systems remain near chance. The same signals also serve as an efficient best-of-N verifier that raises the physical alignment score of a 24-billion-parameter video generator from 50.01 percent to 64.50 percent at reduced compute cost. The work therefore shows that temporal consistency checks can be performed by reading geometry already present in standard image encoders.

Core claim

Indicators of physical plausibility are implicitly captured by five geometric properties of the per-frame embeddings produced by frozen image encoders. In aggregate these properties, called GEOPHYS, correlate with human EEG responses to object-permanence violations, discriminate physically implausible videos from realistic ones at state-of-the-art rates, and function as a low-cost verifier that improves physical alignment during video generation.

What carries the argument

GEOPHYS: the aggregate of five geometric properties of per-frame embeddings from frozen image encoders.

If this is right

  • GEOPHYS separates implausible from realistic videos at 98.3 percent on LikePhys and 93.3 percent on IntPhys2.
  • The same signals outperform V-JEPA 2, GPT-4o, Gemini, and twelve modern video diffusion models on physics-violation detection.
  • When used as a best-of-N verifier, GEOPHYS raises MAGI-1 24B performance on PhysicsIQ from 50.01 percent to 64.50 percent.
  • Physical plausibility assessment becomes possible from emergent geometry in temporal features of image encoders without ad-hoc training.

Where Pith is reading between the lines

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

  • If the geometric signals prove stable across domains, they could support lightweight physics filters inside real-time video pipelines.
  • The reported EEG correlation invites direct comparison between encoder geometry and human perceptual timing on matched stimuli.
  • Video generators might incorporate the five properties as an internal consistency loss rather than relying on external verifiers.
  • New benchmarks that isolate each geometric property could map which aspect tracks which class of physical violation.

Load-bearing premise

The five geometric properties of per-frame embeddings from frozen image encoders implicitly capture indicators of physical plausibility without requiring task-specific training or external models.

What would settle it

Videos containing clear physical violations in which the five geometric properties show no measurable difference from those in matched realistic videos would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.20707 by Alexander Pondaven, Barbara Hammer, Christian Intern\`o, David Klindt, Eero P. Simoncelli, Fabio Pizzati, Francesco Pinto, Habon Issa, Ivan Laptev, Markus Olhofer, Philip Torr.

Figure 1
Figure 1. Figure 1: GEOPHYS signals of plausible vs. violated dynamics in frozen feature space. Paired counterfactuals from LikePhys [1], rendered in Blender [64]. A frozen backbone maps each frame xt to a pooled feature z¯t, yielding a trajectory in representation space (sketched in 2D). Plausible videos produce smooth trajectories; Violated (no momentum conservation) ones show irregular. (1) speed, (2) curvature, (3) predic… view at source ↗
Figure 2
Figure 2. Figure 2: GEOPHYS pipeline. A single video V = (x1, . . . , xT ) feeds each backbone fθ unchanged. Each backbone reads out at the layer ℓ ⋆ selected on a held-out validation split (Appendix B); the per￾frame embedding z¯t = pool f (ℓ ⋆ ) θ (xt)  is spatial-pooled and stacked across time into the trajectory Γ(V) = (z¯1, . . . , z¯T ), on which the geometric signals of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean curvature across layers Blue: plausible. Red: violated (65 video pairs). Violated lies above plausible at every layer of every backbone (p < 10−8 , t-test). Green band: readout layer. Backbone selection. Hypothesis 1 originates in the V1-straightening literature [27, 28], so testing it requires backbones that span V1-like and non-V1-like representations. We use four frozen backbones in two architectur… view at source ↗
Figure 4
Figure 4. Figure 4: Model and brain comparison. (a) Example VOE stimuli for the Create scenario: in valid videos, one object enters and exits occlusion; in invalid videos, one enters but two exit. (b–c) Valid and invalid Create signals from GEOPHYS CORnet-S IT speed (b) and EEG contralateral delay activity (from [6]; c). Both are elevated for the invalid condition after occlusion offset (vertical dashed line). (d–e) Mean vali… view at source ↗
Figure 5
Figure 5. Figure 5: Backbones are complementary. (a) Jaccard overlap of correctly-flagged sets on LikePhys. (b) Greedy stepwise ensemble reaches 98.3% on LikePhys and 93.3% on IntPhys2. (c) Each backbone wins a different physics domain of LikePhys. (d) Each backbone wins a different condition in IntPhys2. Dashed: best baseline (Hunyuan T2V [45] in c, V-JEPA 2 [11] in d). Setup. LikePhys [1] and IntPhys2 [2] present matched pa… view at source ↗
Figure 6
Figure 6. Figure 6: ROC for physics violation detec￾tion. LikePhys (top) and IntPhys2 (bottom). Findings: LikePhys [1] ranks matched pairs via video diffusion mod￾els’ (DMs) likelihood preference: a DM judges plausibility by assigning higher likelihood to the plausible video. All 12 DMs evaluated score 39–56% (near chance, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: PhysicsIQ benchmark and GEOPHYS performance. (A) The five PhysicsIQ scenario categories. (B) Best-of-N (N=16) PhysicsIQ score for five verifiers on five generators; dashed strokes: no-verifier baseline, solid: oracle upper bound. 1 2 4 8 16 Compute Budget (N) 50 55 60 65 70 75 PhysicsIQ Score Random VideoMAE Qwen2.5-VL Qwen3-VL WMReward GeoPhys (Ours) Oracle (UB) [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Test-time scaling on MAGI-1 24B (V2V). PhysicsIQ score as a function of the candidate budget N ∈ {1, . . . , 16}. GEO￾PHYS scales closest to oracle. Setup. We follow the test-time scaling setup of [48]: a generator samples candidates, a verifier ranks them, and an oracle bounds the achievable ceiling. We test all three on PhysicsIQ [3], which provides 198 scenarios, each with a 3 s condition￾ing video and … view at source ↗
Figure 9
Figure 9. Figure 9: PhysicsIQ distributions on V2V (MAGI-1 24B, N=16). Each point is one scenario; the coloured bar is the median, the black the mean, the shaded band the IQR, and triangles are outliers. GEOPHYS shifts the whole distribution towards the Oracle ceiling, not only the mean. reconstruction error on masked spatiotemporal patches as a surprise score. Qwen2.5-VL(BoN) and Qwen3-VL(BoN) [49] prompt a multimodal-LLM wi… view at source ↗
Figure 10
Figure 10. Figure 10: Video-quality metrics (MAGI-1 24B, V2V, N=16). Each metric against the PhysicsIQ score; bars are scenario-level standard errors. Full ± SE values in [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: GEOPHYS inference time per video (N=16, single H100). Compute footprint. GEOPHYS’s verifier path scores at 0.25 s/video and 1.2 GB VRAM with a single backbone (DINOv3 ViT-L), and 1.0 s/video and 2.0 GB with the en￾semble ( [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Cohen’s d of the perceptual-straightening effect across normalised layer depth. Each point is the paired Cohen’s d between turning angles of violated and plausible videos, averaged over 650 LikePhys pairs, at one layer of one backbone. All 57 combinations give d > 0 (p < 10−8 ). DINOv2 rises monotonically; CORnet-S peaks at V1 (d=0.58); DINOv3 peaks mid-depth; VOneNet is flat. B Layer sweep analysis Backb… view at source ↗
Figure 13
Figure 13. Figure 13: Layer sweep on LikePhys (top) and IntPhys2 (bottom). For each backbone, we report per-layer signal accuracy and select the readout layer used in the main text. LikePhys favours deeper layers for ViTs and IT for CORnet-S; IntPhys2 favours mid-level (and V1-like) features. C Additional VOE results In [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Detecting object permanence violations in models and brains: Vanish condition (a) Example VOE stimuli for the Vanish scenario: in valid videos, two objects enter and exit occlusion; in the invalid videos, two objects enter but one exits. (b–c) valid and invalid Vanish signals from GEOPHYS CORnet-S IT (speed) (b) and EEG contralateral delay activity (from [6]; c). Both GEOPHYS and brain signals are elevate… view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative best-of-N selection across physics families (PhysicsIQ V2V, MAGI-1 24B, N=16). Columns left of the dashed line are the shared real conditioning; columns to its right are the continuations, with the real take (Ground truth) for reference [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
read the original abstract

While humans can identify physically implausible events within milliseconds, machine learning approaches addressing the same problem are extremely slow and expensive. They either rely on external multimodal-LLM judges or require ad-hoc modifications to the training procedure. In this work, we argue that indicators of physical plausibility are implicitly captured by five geometric properties of the per-frame embeddings produced by frozen image encoders. In aggregate, we call them GEOPHYS. First, we show that these signals correlate with human EEG responses to two forms of object-permanence violations. Second, GEOPHYS robustly discriminates physically implausible videos from realistic ones, achieving state-of-the-art physics-violation detection: 98.3% on LikePhys and 93.3% on IntPhys2, whereas V-JEPA 2, GPT-4o, Gemini, and twelve modern video diffusion models perform near chance. Third, used as a best-of-N verifier for physical alignment during video generation, GEOPHYS lifts MAGI-1 24B from 50.01% to 64.50% on PhysicsIQ at 1.5x lower wall-clock and 4.65x lower memory than the V-JEPA 2 world-model verifier. Ultimately, GEOPHYS demonstrates that physical plausibility in videos can be assessed by leveraging the emergent geometric properties of temporal features extracted from image encoders.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes GEOPHYS, consisting of five geometric properties of per-frame embeddings from frozen image encoders, which are argued to implicitly capture indicators of physical plausibility without task-specific training. Evidence includes correlation with human EEG responses to object-permanence violations, state-of-the-art discrimination of implausible videos (98.3% on LikePhys, 93.3% on IntPhys2) where baselines including V-JEPA 2, GPT-4o, Gemini, and twelve video diffusion models perform near chance, and improved best-of-N verification for video generation (lifting MAGI-1 from 50.01% to 64.50% on PhysicsIQ at reduced compute).

Significance. If the central claim holds, the result is significant: it shows physical plausibility detection can leverage emergent geometric signals in standard frozen encoders, avoiding expensive LLM judges or ad-hoc training. Credit is due for the three independent lines of evidence (EEG correlation, benchmark discrimination, and generation verifier), the parameter-free nature, and the efficiency gains (1.5x lower wall-clock, 4.65x lower memory than V-JEPA 2). This could impact video understanding and generation pipelines.

major comments (2)
  1. [Methods] The abstract and summary claim the five properties are defined without free parameters or fitting, but the exact definitions, formulas, and aggregation method for GEOPHYS must be specified with equations in the methods section to allow verification that they are independently derived and not circular.
  2. [Experiments] Table or section reporting the 98.3% / 93.3% accuracies: the evaluation protocol (e.g., how per-frame embeddings are processed into video-level scores, dataset splits, and whether any hyperparameter tuning occurred) is load-bearing for the SOTA claim and must be detailed with controls for dataset bias.
minor comments (2)
  1. [Abstract] The abstract refers to 'five geometric properties' without naming them (e.g., what specific geometric measures like distances, angles, or curvatures); adding names or a brief list would improve clarity.
  2. [EEG Experiments] The EEG correlation experiment should include the number of subjects, trial counts, and statistical measures (e.g., correlation coefficients with p-values) for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and for recognizing the potential significance of GEOPHYS. We address the two major comments below and will incorporate the requested clarifications into the revised manuscript.

read point-by-point responses

  1. Referee: [Methods] The abstract and summary claim the five properties are defined without free parameters or fitting, but the exact definitions, formulas, and aggregation method for GEOPHYS must be specified with equations in the methods section to allow verification that they are independently derived and not circular.

    Authors: We agree that explicit equations are necessary for full verification. The five properties are defined from first principles on the geometry of the embedding manifold (trajectory curvature, pairwise cosine distances, volume of the convex hull of consecutive embeddings, eigenvalue spread of the temporal covariance, and a normalized displacement norm), with no learned parameters or data-dependent thresholds. In the revision we will add a dedicated Methods subsection containing the precise mathematical definitions, the closed-form aggregation into the scalar GEOPHYS score, and a short proof sketch confirming independence from any fitting procedure. revision: yes

  2. Referee: [Experiments] Table or section reporting the 98.3% / 93.3% accuracies: the evaluation protocol (e.g., how per-frame embeddings are processed into video-level scores, dataset splits, and whether any hyperparameter tuning occurred) is load-bearing for the SOTA claim and must be detailed with controls for dataset bias.

    Authors: The current manuscript already states that no hyper-parameters are tuned and that the same fixed geometric thresholds are used across all datasets, but we accept that a more explicit protocol description is warranted. In the revision we will expand the Experiments section with: (i) the exact per-frame to video-level aggregation rule, (ii) the precise train/test splits employed for LikePhys and IntPhys2, (iii) confirmation that zero hyper-parameter search was performed, and (iv) additional bias controls (shuffled-label baselines and cross-dataset transfer results). These additions will strengthen the SOTA claim without altering any reported numbers. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper presents five geometric properties of per-frame embeddings from frozen image encoders as emergent indicators of physical plausibility. Validation proceeds via independent empirical lines: correlation with human EEG responses to object-permanence violations, 98.3% and 93.3% accuracy on LikePhys and IntPhys2 (where strong baselines fail near chance), and measurable gains as a best-of-N verifier on PhysicsIQ. No equations, definitions, or self-citations in the provided text reduce any claimed result to a fitted input, self-defined quantity, or author-prior ansatz by construction. The approach is self-contained against external benchmarks without load-bearing internal fitting or renaming.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract does not specify any free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5810 in / 976 out tokens · 59189 ms · 2026-06-27T03:32:15.915632+00:00 · methodology

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

92 extracted references · 4 canonical work pages

  1. [1]

    LikePhys: Evaluating intuitive physics understanding in video diffusion models via likelihood preference.arXiv preprint arXiv:2510.11512, 2025

    Jianhao Yuan, Fabio Pizzati, Francesco Pinto, Lars Kunze, Ivan Laptev, Paul Newman, Philip Torr, and Daniele De Martini. LikePhys: Evaluating intuitive physics understanding in video diffusion models via likelihood preference.arXiv preprint arXiv:2510.11512, 2025. (Cited on pages 1, 2, 2, 2, 3, 5, 5, 5, 6, 6, 7, 7, 7, 8, and 21.)

    arXiv 2025

  2. [2]

    IntPhys 2: Benchmarking intuitive physics understanding in complex synthetic environments.arXiv preprint arXiv:2506.09849, 2025

    Florian Bordes, Quentin Garrido, Justine T Kao, Adina Williams, Michael Rabbat, and Em- manuel Dupoux. IntPhys 2: Benchmarking intuitive physics understanding in complex synthetic environments.arXiv preprint arXiv:2506.09849, 2025. (Cited on pages 1, 2, 2, 2, 3, 5, 6, 6, 7, 7, 7, 7, 8, 18, 21, and 22.)

    arXiv 2025

  3. [3]

    Do generative video models understand physical principles?, 2025

    Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models understand physical principles?, 2025. URL https://arxiv.org/abs/2501. 09038. (Cited on pages 1, 2, 2, 3, 5, 8, and 22.)

    2025

  4. [4]

    The violation-of-expectation paradigm: A conceptual overview.Psychological Review, 131(3):716, 2024

    Francesco Margoni, Luca Surian, and Renée Baillargeon. The violation-of-expectation paradigm: A conceptual overview.Psychological Review, 131(3):716, 2024. (Cited on pages 1 and 5.)

    2024

  5. [5]

    Investigating looking and social looking measures as an index of infant violation of expectation.Developmental Science, 20(6):e12452, 2017

    Kirsty Dunn and J Gavin Bremner. Investigating looking and social looking measures as an index of infant violation of expectation.Developmental Science, 20(6):e12452, 2017. (Cited on pages 1 and 5.)

    2017

  6. [6]

    Electrophysiology reveals that intuitive physics guides visual tracking and working memory.Open Mind, 8: 1425–1446, 2024

    Halely Balaban, Kevin A Smith, Joshua B Tenenbaum, and Tomer D Ullman. Electrophysiology reveals that intuitive physics guides visual tracking and working memory.Open Mind, 8: 1425–1446, 2024. (Cited on pages 1, 2, 5, 5, 5, 6, 6, 6, 6, 10, 20, and 21.)

    2024

  7. [7]

    Violations of physical and psycholog- ical expectations in the human adult brain.Imaging Neuroscience, 2:imag–2–00068, 02 2024

    Shari Liu, Kirsten Lydic, Lingjie Mei, and Rebecca Saxe. Violations of physical and psycholog- ical expectations in the human adult brain.Imaging Neuroscience, 2:imag–2–00068, 02 2024. ISSN 2837-6056. doi: 10.1162/imag_a_00068. URL https://doi.org/10.1162/imag_a_ 00068. (Cited on pages 1, 2, 2, 2, 5, 10, and 21.)

    work page doi:10.1162/imag_a_00068 2024

  8. [8]

    Intuitive physics understanding emerges from self-supervised pretraining on natural videos, 2025

    Quentin Garrido, Nicolas Ballas, Mahmoud Assran, Adrien Bardes, Laurent Najman, Michael Rabbat, Emmanuel Dupoux, and Yann LeCun. Intuitive physics understanding emerges from self-supervised pretraining on natural videos, 2025. URL https://arxiv.org/abs/2502. 11831. (Cited on pages 1 and 3.)

    2025

  9. [9]

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

    Peiyao Wang, Weining Wang, and Qi Li. Physcorr: Dual-reward dpo for physics-constrained text-to-video generation with automated preference selection, 2025. URL https://arxiv. org/abs/2511.03997. (Cited on page 1.)

    arXiv 2025

  10. [10]

    Inference-time physics alignment of video generative models with latent world models

    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 generative models with latent world models. arXiv preprint arXiv:2601.10553, 2026. (Cited on pages 1, 2, 2, 9, 9, 9, 9, and 26.) G...

    arXiv 2026

  11. [11]

    V-jepa 2: Self-supervised video models enable understanding, prediction and planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba, Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, Xia...

  12. [12]

    (Cited on pages 1, 2, 2, 2, 6, 7, 7, and 9.)

    URL https://arxiv.org/abs/2506.09985. (Cited on pages 1, 2, 2, 2, 6, 7, 7, and 9.)

    Pith/arXiv arXiv

  13. [13]

    Travl: A recipe for making video-language models better judges of physics implausibility, 2025

    Saman Motamed, Minghao Chen, Luc Van Gool, and Iro Laina. Travl: A recipe for making video-language models better judges of physics implausibility, 2025. URL https://arxiv. org/abs/2510.07550. (Cited on pages 1 and 2.)

    arXiv 2025

  14. [14]

    Tenenbaum, and Tomer D

    Kevin Smith, Lingjie Mei, Shunyu Yao, Jiajun Wu, Elizabeth Spelke, Joshua B. Tenenbaum, and Tomer D. Ullman. Modeling expectation violation in intuitive physics with coarse probabilistic object representations. InAdvances in Neural Information Processing Systems (NeurIPS), volume 32, 2019. (Cited on pages 1 and 5.)

    2019

  15. [15]

    Interaction networks for learning about objects, relations and physics

    Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, and Koray kavukcuoglu. Interaction networks for learning about objects, relations and physics. InProceed- ings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 4509–4517, Red Hook, NY , USA, 2016. Curran Associates Inc. ISBN 9781510838819. (Ci...

    2016

  16. [16]

    Galileo: Perceiving physical object properties by integrating a physics engine with deep learn- ing

    Jiajun Wu, Ilker Yildirim, Joseph J Lim, Bill Freeman, and Josh Tenenbaum. Galileo: Perceiving physical object properties by integrating a physics engine with deep learn- ing. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. URL https://pro...

    2015

  17. [17]

    Lighting (in)consistency of paint by text, 2022

    Hany Farid. Lighting (in)consistency of paint by text, 2022. URL https://arxiv.org/abs/ 2207.13744. (Cited on page 1.)

    arXiv 2022

  18. [18]

    O’brien, and Hany Farid

    Eric Kee, James F. O’brien, and Hany Farid. Exposing photo manipulation from shading and shadows.ACM Trans. Graph., 33(5), September 2014. ISSN 0730-0301. doi: 10.1145/2629646. URLhttps://doi.org/10.1145/2629646. (Cited on page 1.)

    work page doi:10.1145/2629646 2014

  19. [19]

    Distinguishing authentic from ai-generated explosions using spatiotemporal dynamics

    Sarah Barrington and Hany Farid. Distinguishing authentic from ai-generated explosions using spatiotemporal dynamics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 10659–10667, June 2026. (Cited on page 1.)

    2026

  20. [20]

    Perspective (in)consistency of paint by text, 2022

    Hany Farid. Perspective (in)consistency of paint by text, 2022. URL https://arxiv.org/ abs/2206.14617. (Cited on page 1.)

    arXiv 2022

  21. [21]

    Learning invariance from transformation sequences.Neural Computation, 3(2): 194–200, 1991

    Peter Földiák. Learning invariance from transformation sequences.Neural Computation, 3(2): 194–200, 1991. (Cited on page 2 and 2.)

    1991

  22. [22]

    Sejnowski

    Laurenz Wiskott and Terrence J. Sejnowski. Slow feature analysis: Unsupervised learning of invariances.Neural Computation, 14(4):715–770, 04 2002. ISSN 0899-7667. doi: 10.1162/ 089976602317318938. URL https://doi.org/10.1162/089976602317318938. (Cited on page 2 and 2.)

    work page doi:10.1162/089976602317318938 2002

  23. [23]

    Tangent prop: A formalism for specifying selected invariances in an adaptive network

    Patrice Simard, Bernard Victorri, Yann LeCun, and John Denker. Tangent prop: A formalism for specifying selected invariances in an adaptive network. InAdvances in Neural Information Processing Systems, volume 4, pages 895–903, 1991. (Cited on page 2 and 2.)

    1991

  24. [24]

    Learning lie groups for invariant visual perception

    Rajesh Rao and Daniel Ruderman. Learning lie groups for invariant visual perception. In M. Kearns, S. Solla, and D. Cohn, editors,Advances in Neural Information Processing Systems, volume 11. MIT Press, 1998. URL https://proceedings.neurips.cc/paper_files/ paper/1998/file/277281aada22045c03945dcb2ca6f2ec-Paper.pdf. (Cited on page 2 and 2.) GEOPHYS: The Ge...

    1998

  25. [25]

    Learning to linearize under uncertainty

    Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty

  26. [26]

    (Cited on pages 2, 2, and 4.)

    URLhttps://arxiv.org/abs/1506.03011. (Cited on pages 2, 2, and 4.)

    Pith/arXiv arXiv

  27. [27]

    Hénaff and Eero P

    Olivier J. Hénaff and Eero P. Simoncelli. Geodesics of learned representations.CoRR, abs/1511.06394, 2015. URL https://api.semanticscholar.org/CorpusID:2208884. (Cited on page 2.)

    arXiv 2015

  28. [28]

    Learning predictable and robust neural representations by straightening image sequences.Advances in Neural Information Processing Systems, 37:40316–40335, 2024

    Xueyan Niu, Cristina Savin, and Eero P Simoncelli. Learning predictable and robust neural representations by straightening image sequences.Advances in Neural Information Processing Systems, 37:40316–40335, 2024. (Cited on page 2 and 2.)

    2024

  29. [29]

    Perceptual straightening of natural videos.Nature Neuroscience, 22(6):984–991, 2019

    Olivier J Hénaff, Robbe LT Goris, and Eero P Simoncelli. Perceptual straightening of natural videos.Nature Neuroscience, 22(6):984–991, 2019. (Cited on pages 2, 3, 5, 18, and 18.)

    2019

  30. [30]

    Primary visual cortex straightens natural video trajectories

    Olivier J Hénaff, Yoon Bai, Julie A Charlton, Ian Nauhaus, Eero P Simoncelli, and Robbe LT Goris. Primary visual cortex straightens natural video trajectories. InNature Communications,

  31. [31]

    (Cited on pages 2, 2, 3, 5, and 18.)

  32. [32]

    Anne Harrington, Vasha DuTell, Ayush Tewari, Mark Hamilton, Simon Stent, Ruth Rosenholtz, and William T. Freeman. Exploring perceptual straightness in learned visual representations. InThe Eleventh International Conference on Learning Representations, 2023. URL https: //openreview.net/forum?id=4cOfD2qL6T. (Cited on page 2 and 2.)

    2023

  33. [33]

    Zonneveld, Pascal Mettes, and Iris Groen

    Anne W. Zonneveld, Pascal Mettes, and Iris Groen. Straightening of natural visual sequences in video DNNs: the role of locality and temporal coherence. InCognitive Computational Neuroscience (CCN), Amsterdam, The Netherlands, 2025. URL https://2025.ccneuro. org/poster/?id=EJMZ9Os8jG. (Cited on pages 2, 2, and 3.)

    2025

  34. [34]

    Chang, Ashesh Rambachan, and Sendhil Mullainathan

    Keyon Vafa, Peter G. Chang, Ashesh Rambachan, and Sendhil Mullainathan. What has a foun- dation model found? inductive bias reveals world models. InForty-second International Confer- ence on Machine Learning, 2025. URLhttps://openreview.net/forum?id=i9npQatSev. (Cited on pages 2 and 11.)

    2025

  35. [35]

    The observer effect in world models: Invasive adaptation corrupts latent physics.arXiv preprint arXiv:2602.12218, 2026

    Christian Internò, Jumpei Yamaguchi, Loren Amdahl-Culleton, Markus Olhofer, David Klindt, and Barbara Hammer. The observer effect in world models: Invasive adaptation corrupts latent physics.arXiv preprint arXiv:2602.12218, 2026. (Cited on pages 2 and 11.)

    arXiv 2026

  36. [36]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick L...

    2024

  37. [37]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    Pith/arXiv arXiv 2025

  38. [38]

    Brain-like object recognition with high-performing shallow recurrent ANNs

    Jonas Kubilius, Martin Schrimpf, Kohitij Kar, Rishi Rajalingham, Ha Hong, Najib Majaj, Elias Issa, Pouya Bashivan, Jonathan Prescott-Roy, Kailyn Schmidt, et al. Brain-like object recognition with high-performing shallow recurrent ANNs. InAdvances in Neural Information Processing Systems, 2019. (Cited on pages 2, 5, 7, 7, and 18.)

    2019

  39. [39]

    Simulating a primary visual cortex at the front of CNNs improves robustness to image perturbations

    Joel Dapello, Tiago Marques, Martin Schrimpf, Franziska Geiger, David Cox, and James J DiCarlo. Simulating a primary visual cortex at the front of CNNs improves robustness to image perturbations. InAdvances in Neural Information Processing Systems, 2020. (Cited on pages 2, 3, 5, 7, 7, and 19.) GEOPHYS: The Geometry of Physical Plausibility 14

    2020

  40. [40]

    Eeg decoding reveals neural predictions for naturalistic material behaviors.Journal of Neuroscience, 43(29):5406–5413, 2023

    Daniel Kaiser, Rico Stecher, and Katja Doerschner. Eeg decoding reveals neural predictions for naturalistic material behaviors.Journal of Neuroscience, 43(29):5406–5413, 2023. (Cited on pages 2, 5, and 21.)

    2023

  41. [41]

    Gpt-4 technical report, 2024

    OpenAI. Gpt-4 technical report, 2024. URL https://arxiv.org/abs/2303.08774. (Cited on pages 2, 2, and 7.)

    Pith/arXiv arXiv 2024

  42. [42]

    Gemini: A family of highly capable multimodal models, 2025

    Gemini Team. Gemini: A family of highly capable multimodal models, 2025. URL https: //arxiv.org/abs/2312.11805. (Cited on pages 2, 2, 7, and 7.)

    Pith/arXiv arXiv 2025

  43. [43]

    MAGI-1: Autoregressive video generation at scale.arXiv preprint arXiv:2505.13211,

    Sand AI. MAGI-1: Autoregressive video generation at scale.arXiv preprint arXiv:2505.13211,

    Pith/arXiv arXiv

  44. [44]

    (Cited on pages 2 and 8.)

  45. [45]

    Bear, Elias Wang, Damian Mrowca, Felix J

    Daniel M. Bear, Elias Wang, Damian Mrowca, Felix J. Binder, Hsiao-Yu Fish Tung, R. T. Pramod, Cameron Holdaway, Sirui Tao, Kevin Smith, Fan-Yun Sun, Li Fei-Fei, Nancy Kanwisher, Joshua B. Tenenbaum, Daniel L. K. Yamins, and Judith E. Fan. Physion: Evaluating physical prediction from vision in humans and machines, 2022. URL https: //arxiv.org/abs/2106.0826...

    arXiv 2022

  46. [46]

    IntPhys: A benchmark for visual intuitive physics reasoning

    Ronan Riochet, Mario Ynocente Castro, Mathieu Bernard, Adam Lerer, Rob Fergus, Véronique Izard, and Emmanuel Dupoux. IntPhys: A benchmark for visual intuitive physics reasoning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5016–5025, 2022. (Cited on page 2.)

    2022

  47. [47]

    Grasp: A novel benchmark for evaluating language grounding and situated physics understanding in multimodal language models

    Serwan Jassim, Mario Holubar, Annika Richter, Cornelius Wolff, Xenia Ohmer, and Elia Bruni. Grasp: A novel benchmark for evaluating language grounding and situated physics understanding in multimodal language models. 2024. URL https://arxiv.org/abs/2311. 09048. (Cited on page 2.)

    2024

  48. [48]

    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, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, T...

    Pith/arXiv arXiv 2025

  49. [49]

    HunyuanVideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, et al. HunyuanVideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. (Cited on pages 2, 6, and 7.)

    Pith/arXiv arXiv 2024

  50. [50]

    CogVideoX: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    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.arXiv preprint arXiv:2408.06072, 2024. (Cited on pages 2, 7, 7, 7, and 8.)

    Pith/arXiv arXiv 2024

  51. [51]

    VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=AhccnBXSne. (Cited on pages 2, 2, 7, and 8.)

    2022

  52. [52]

    Video- t1: Test-time scaling for video generation.arXiv preprint arXiv:2503.18942, 2025

    Fangfu Liu, Hanyang Wang, Yimo Cai, Kaiyan Zhang, Xiaohang Zhan, and Yueqi Duan. Video- t1: Test-time scaling for video generation.arXiv preprint arXiv:2503.18942, 2025. (Cited on pages 2 and 8.)

    arXiv 2025

  53. [53]

    Qwen3 technical report, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin GEOPHYS: The Geometry of Physical Plausibility 15 Yang, Jiaxi Yang, Jing Zhou, Jingren Zh...

    Pith/arXiv arXiv 2025

  54. [54]

    Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation

    Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, Kai Wang, Quy Duc Do, Yuansheng Ni, Bohan Lyu, Yaswanth Narsupalli, Rongqi Fan, Zhiheng Lyu, Yuchen Lin, and Wenhu Chen. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. In Procee...

    2024

  55. [55]

    Videophy: Evaluating physical commonsense for video generation.arXiv preprint arXiv:2406.03520, 2024

    Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation.arXiv preprint arXiv:2406.03520, 2024. (Cited on page 2.)

    Pith/arXiv arXiv 2024

  56. [56]

    Vigor: Video geometry-oriented reward for temporal generative alignment, 2026

    Tengjiao Yin, Jinglei Shi, Heng Guo, and Xi Wang. Vigor: Video geometry-oriented reward for temporal generative alignment, 2026. URL https://arxiv.org/abs/2603.16271. (Cited on page 2.)

    arXiv 2026

  57. [57]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Pith/arXiv arXiv

  58. [58]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InProceedings of ICLR, 2024. (Cited on page 2.)

    2024

  59. [59]

    Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314,

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314,

    Pith/arXiv arXiv

  60. [60]

    (Cited on pages 2 and 10.)

  61. [61]

    Le, Christopher Ré, and Azalia Mirhoseini

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V . Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024. (Cited on pages 2 and 10.)

    Pith/arXiv arXiv 2024

  62. [62]

    A general framework for inference-time scaling and steering of diffusion models

    Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models. InForty-second International Conference on Machine Learning, 2025. URL https: //openreview.net/forum?id=Jp988ELppQ. (Cited on pages 2 and 10.)

    2025

  63. [63]

    Inference-time text-to- video alignment with diffusion latent beam search

    Yuta Oshima, Masahiro Suzuki, Yutaka Matsuo, and Hiroki Furuta. Inference-time text-to- video alignment with diffusion latent beam search. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?id= c9EAmyYPOv. (Cited on page 2.)

    2025

  64. [64]

    Verifier matters: Enhancing inference-time scaling for video diffusion models

    Lorenzo Baraldi, Davide Bucciarelli, Zifan Zeng, Chongzhe Zhang, Qunli Zhang, Marcella Cornia, Lorenzo Baraldi, Feng Liu, Zheng Hu, and Rita Cucchiara. Verifier matters: Enhancing inference-time scaling for video diffusion models. In36th British Machine Vision Confer- ence 2025, BMVC 2025, Sheffield, UK, November 24-27, 2025. BMV A, 2025. URL https: //bmv...

    2025

  65. [65]

    Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects.Nature Neuroscience, 2(1):79–87,

    Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects.Nature Neuroscience, 2(1):79–87,

  66. [66]

    The free-energy principle: a unified brain theory?Nature Reviews Neuroscience, 11(2):127–138, 2010

    Karl Friston. The free-energy principle: a unified brain theory?Nature Reviews Neuroscience, 11(2):127–138, 2010. (Cited on page 2.) GEOPHYS: The Geometry of Physical Plausibility 16

    2010

  67. [67]

    Ai-generated video detection via perceptual straightening

    Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, and David Klindt. Ai-generated video detection via perceptual straightening. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors,Advances in Neural Information Processing Systems, volume 38, pages 20672–20705. Curran Asso- ciates, Inc., 20...

    2025

  68. [68]

    Grab-3d: Detecting ai-generated videos from 3d geometric temporal consistency, 2025

    Wenhan Chen, Sezer Karaoglu, and Theo Gevers. Grab-3d: Detecting ai-generated videos from 3d geometric temporal consistency, 2025. URLhttps://arxiv.org/abs/2512.13665. (Cited on page 3.)

    arXiv 2025

  69. [69]

    Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018

    Blender Online Community.Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. URL http://www.blender. org. (Cited on pages 3 and 7.)

    2018

  70. [70]

    Video diffusion models: A survey, 2024

    Andrew Melnik, Michal Ljubljanac, Cong Lu, Qi Yan, Weiming Ren, and Helge Ritter. Video diffusion models: A survey, 2024. URL https://arxiv.org/abs/2405.03150. (Cited on page 3.)

    arXiv 2024

  71. [71]

    Issa, and James J

    Kohitij Kar, Jonas Kubilius, Kailyn Schmidt, Elias B. Issa, and James J. DiCarlo. Evidence that recurrent circuits are critical to the ventral stream’s execution of core object recognition behavior.Nature Neuroscience, 22(6):974–983, 2019. (Cited on page 5.)

    2019

  72. [72]

    David J. Heeger. Normalization of cell responses in cat striate cortex.Visual Neuroscience, 9 (2):181–197, 1992. (Cited on page 5.)

    1992

  73. [73]

    Matteo Carandini and David J. Heeger. Normalization as a canonical neural computation. Nature Reviews Neuroscience, 13(1):51–62, 2012. (Cited on page 5.)

    2012

  74. [74]

    Contralateral delay activity provides a neural measure of the number of representations in visual working memory.Journal of neurophysiology, 103(4):1963–1968, 2010

    Akiko Ikkai, Andrew W McCollough, and Edward K V ogel. Contralateral delay activity provides a neural measure of the number of representations in visual working memory.Journal of neurophysiology, 103(4):1963–1968, 2010. (Cited on page 5.)

    1963

  75. [75]

    Electrophysiological measures of maintaining representations in visual working memory.Cortex, 43(1):77–94, 2007

    Andrew W McCollough, Maro G Machizawa, and Edward K V ogel. Electrophysiological measures of maintaining representations in visual working memory.Cortex, 43(1):77–94, 2007. (Cited on page 5.)

    2007

  76. [76]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=Fx2SbBgcte. (Cited on page 7 and 7.)

    2024

  77. [77]

    Cosmos world foundation model platform for physical ai, 2025

    NVIDIA team. Cosmos world foundation model platform for physical ai, 2025. URL https: //arxiv.org/abs/2501.03575. (Cited on page 7.)

    Pith/arXiv arXiv 2025

  78. [78]

    Modelscope text-to-video technical report, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report, 2023. URL https://arxiv.org/abs/2308. 06571. (Cited on page 7 and 7.)

    2023

  79. [79]

    Mochi 1: A new SOTA in open-source video generation models

    Genmo Team. Mochi 1: A new SOTA in open-source video generation models. https: //github.com/genmoai/mochi, 2024. (Cited on page 7.)

    2024

  80. [80]

    Ltx-video: Realtime video latent diffusion, 2024

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richard- son, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion, 2024. URLhttps://arxiv.org/abs/2501.00103. (Cited on page 7.)

    Pith/arXiv arXiv 2024

Showing first 80 references.