pith. sign in

arxiv: 2606.30544 · v1 · pith:7YAIUFSLnew · submitted 2026-06-29 · 💻 cs.AI

Latent Actions from Factorized Transition Effects under Agent Ambiguity

Heejeong Nam , Chandradithya S Jonnalagadda , Harshit Aggarwal , Eric Xu , Randall Balestriero This is my paper

Pith reviewed 2026-06-30 05:58 UTC · model grok-4.3

classification 💻 cs.AI
keywords latent action modelstransition factorizationunsupervised action learningmulti-object scenespolicy learningzero-shot transferinverse dynamicsvisual ambiguity
0
0 comments X

The pith

Decomposing visual transitions into sparse primitives yields reusable latent actions despite distractors and ambiguity.

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

The paper establishes that standard latent action models fail in multi-object scenes because visual effects entangle agent motion with distractors and camera changes. Observed Transition Factorization addresses this by breaking each transition into a sparse set of primitives that serve as an intermediate interface. These primitives feed into OTF-LAM to produce action-like latents via the usual inverse-forward dynamics setup, and into a decoder-free variant that works in frozen DINOv2 space. The resulting primitives transfer without retraining across carrier and morphology changes, and the downstream policies match or exceed baselines when transition sources are mixed.

Core claim

Observed Transition Factorization (OTF) decomposes each transition into a sparse set of observed transition primitives. Using these primitives as the transition interface, OTF-LAM abstracts motion primitives into action-like latents within the standard inverse-forward dynamics framework, and OTF-LAM-Dino predicts future states in a frozen DINOv2 representation space without a decoder. This structure turns the mixture of agent motion, distractors, and background changes into reusable transition effects from which action-like latents can be formed more robustly.

What carries the argument

Observed Transition Factorization (OTF), the decomposition of each visual transition into a sparse set of observed transition primitives that isolate action sources.

If this is right

  • OTF primitives transfer zeroshot across controlled carrier and morphology shifts.
  • Downstream policy learning matches or outperforms baselines under complex transition ambiguity.
  • The factorization supplies an intermediate representation that structures mixed visual effects into reusable transition effects.
  • Action-like latents become more robust when formed from the factorized primitives rather than raw transitions.

Where Pith is reading between the lines

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

  • The same primitives could be reused across entirely different robot embodiments if the sparsity pattern holds for the new morphology.
  • Replacing the DINOv2 space with other frozen visual encoders might preserve the zero-shot transfer property.
  • The method opens a path to learning from unlabeled video without explicit action labels even when the camera itself moves.
  • Extending the factorization to handle temporal sequences longer than single transitions could improve long-horizon planning.

Load-bearing premise

Visual transition effects in multi-object scenes can be decomposed into a sparse set of observed transition primitives that isolate the underlying action source without supervision.

What would settle it

A controlled test scene where known agent actions are mixed with distractors, in which the extracted OTF primitives fail to separate agent-specific effects and the resulting policies underperform non-factorized baselines.

Figures

Figures reproduced from arXiv: 2606.30544 by Chandradithya S Jonnalagadda, Eric Xu, Harshit Aggarwal, Heejeong Nam, Randall Balestriero.

Figure 1
Figure 1. Figure 1: Motivation for agent ambiguity in observation-only latent action learning. Unlike action-conditioned world models, which receive the actor and action as part of the input, LAMs infer action￾like latents only from visual transitions. (LAMs) (Schmidt & Jiang, 2024; Ye et al., 2025) have been proposed to learn action-like proxies directly from obser￾vation transitions. A typical LAM encodes an observation pai… view at source ↗
Figure 2
Figure 2. Figure 2: From causal generation to inverse inference. Blue arrows denote the causal pathway from action to visual change, while red arrows denote the inverse problem faced by observation-only latent action learning. 2. From Actions to Observed Effects LAMs trained from observation-only transitions do not re￾cover actions directly, but representations of their visible consequences after being mediated by state, embo… view at source ↗
Figure 3
Figure 3. Figure 3: Observed-transition vocabulary learning with a VQ￾VAE. Motion-centered transition inputs are patchified, quantized into a shared codebook of reusable observed-transition primitives, and decoded with reference-frame conditioning to reconstruct the observed-transition. Patchwise vector quantization. Given ot ∈ R C×H×W , we partition it into P non-overlapping spatial patches and encode each patch with a shall… view at source ↗
Figure 4
Figure 4. Figure 4: Latent-action learning on top of the pretrained observed-transition vocabulary. Given the extracted observed-transition factors Et and the current frame xt, the model builds state-aware factor tokens, scores them with a relevance gate, aggregates the selected factors into a compact action-like latent z act t , and uses a reference-conditioned decoder to predict the next observation. contains the informatio… view at source ↗
Figure 6
Figure 6. Figure 6: Policy Training. The latent motion space is distilled into a policy via behavioral cloning (left). An action decoder then maps these primitives to ground-truth actions at, utilizing the pixel frame to associate motions with specific objects (right). environment-specific. OTF-LAM-Dino pushes this separa￾tion further by replacing pixel-space prediction with pre￾diction in a frozen DINOv2 representation space… view at source ↗
Figure 5
Figure 5. Figure 5: OTF-LAM-Dino uses frozen DINOv2 as an encoder and frozen OTF module as a motion primitive extractor. factorizer has identified the controlled agent. It only learns which observed-transition factors are useful for forming a compact latent for prediction. Forward dynamics. We predict the next observation from the current frame and the latent action xˆt+τ = pθ(xt, zact t ). Features from xt provide appearance… view at source ↗
Figure 7
Figure 7. Figure 7: Example of learned OTF codes assigned to the motion input. More visualization results are in Appendix D.2 [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-morphology transfer in DCS using only the pre￾trained observed-transition factorizer. The vocabulary is trained on walker-run and evaluated zero-shot on cheetah-run. 4.3. Reusability of Latent Action Primitives In this section, we evaluate the observed-transition prim￾itives learned by the pretrained OTF tokenizer. A useful motion vocabulary should transfer across visual carriers and embodied morphol… view at source ↗
Figure 9
Figure 9. Figure 9: Controlled carrier transfer on Moving MNIST [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: (a) Bar chart comparing the relative mean return of latent action model variants on the DCS cheetah-run downstream policy evaluation task. (b) Downstream evaluation of OTF-LAM with different motion vocabulary sizes. Error bar is computed as the RMS of standard deviations per seed. environment-specific regularities. This suggests a broader direction for pushing the generalization boundary of latent action … view at source ↗
read the original abstract

Latent Action Models (LAMs) learn action-like proxies from observation transitions. However, in multi-object or distractor-rich scenes, these visual effects mix agent motion with distractors, camera dynamics, and background changes, making the underlying action source ambiguous without supervision. Structuring this mixture as reusable transition effects provides an intermediate representation from which action-like latents can be more robustly formed. We introduce Observed Transition Factorization (OTF), which decomposes each transition into a sparse set of observed transition primitives. Using these primitives as the transition interface, we propose OTF-LAM, which abstracts motion primitives into action-like latents within the standard inverse-forward dynamics framework, and OTF-LAM-Dino, a decoder-free variant that predicts future states in a frozen DINOv2 representation space. Empirically, OTF primitives transfer zeroshot across controlled carrier and morphology shifts, showing reusability. Furthermore, downstream policy learning results match or outperform baselines under complex transition ambiguity.

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 / 0 minor

Summary. The paper introduces Observed Transition Factorization (OTF) to decompose visual observation transitions into a sparse set of observed transition primitives without supervision in multi-object or distractor-rich scenes. These primitives serve as an interface for OTF-LAM, which learns action-like latents via the standard inverse-forward dynamics framework, and for the decoder-free OTF-LAM-Dino variant that operates in frozen DINOv2 space. The central claims are that the resulting primitives transfer zeroshot across controlled carrier and morphology shifts (demonstrating reusability) and that downstream policy learning matches or outperforms baselines under complex transition ambiguity.

Significance. If the empirical claims hold with rigorous validation, the work could meaningfully advance unsupervised learning of reusable latent actions in visually ambiguous settings, a persistent challenge in visual RL and robotics. The factorization into sparse primitives offers a structured intermediate representation that may reduce the impact of distractors, and the decoder-free DINOv2 variant is a practical design choice. No machine-checked proofs or parameter-free derivations are present.

major comments (2)
  1. [Abstract] Abstract: The abstract asserts empirical support for zeroshot transfer of OTF primitives and competitive downstream policy performance, but the manuscript supplies no experimental details, metrics, baselines, datasets, or implementation descriptions, preventing any evaluation of the support for these central claims.
  2. [Method] Method section (OTF definition): No derivation or analysis is provided showing why the sparsity constraint or factorization objective isolates the underlying agent action source from distractors, camera effects, or background changes rather than capturing spurious correlations; without this, the claimed reusability and policy gains do not necessarily follow from the stated mechanism.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review. We address the two major comments point-by-point below, clarifying the role of the abstract and providing additional motivation for the OTF factorization while remaining faithful to the manuscript's content.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The abstract asserts empirical support for zeroshot transfer of OTF primitives and competitive downstream policy performance, but the manuscript supplies no experimental details, metrics, baselines, datasets, or implementation descriptions, preventing any evaluation of the support for these central claims.

    Authors: Abstracts are concise summaries by design and do not contain experimental details; the manuscript's Experiments section provides full descriptions of the environments, datasets (including multi-object and distractor-rich scenes), metrics (zeroshot transfer success and policy returns), baselines (standard LAMs and variants), and implementation (training objectives, architectures for OTF-LAM and OTF-LAM-Dino). The abstract's claims are directly supported by those results. We can add a brief parenthetical reference to the experimental protocol if the editor prefers. revision: partial

  2. Referee: [Method] Method section (OTF definition): No derivation or analysis is provided showing why the sparsity constraint or factorization objective isolates the underlying agent action source from distractors, camera effects, or background changes rather than capturing spurious correlations; without this, the claimed reusability and policy gains do not necessarily follow from the stated mechanism.

    Authors: The OTF objective combines a reconstruction loss with an explicit sparsity penalty on the transition primitives, which is motivated by the assumption that agent-induced effects are sparse relative to background or distractor motion. This inductive bias is validated empirically through controlled carrier/morphology transfer experiments and qualitative factorization visualizations that separate agent motion from distractors. While a formal proof of isolation is absent, the design follows standard sparse coding principles used in disentanglement literature, and the downstream policy gains are measured directly against non-factorized baselines under identical ambiguity conditions. We can add a short paragraph in the Method section elaborating on these inductive biases. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper proposes Observed Transition Factorization (OTF) as a novel unsupervised decomposition of visual transitions into sparse primitives, then integrates these primitives into the existing inverse-forward dynamics framework to form OTF-LAM and OTF-LAM-Dino. All performance claims (zeroshot transfer across carrier/morphology shifts and downstream policy results) are framed as empirical outcomes rather than closed-form derivations. No equations, self-citations, or fitted parameters are shown in the abstract or description that reduce the reusability or isolation claims to quantities defined by construction within the same work. The derivation chain remains self-contained against external benchmarks and does not exhibit self-definitional, fitted-input, or self-citation load-bearing patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The abstract introduces OTF as a new decomposition without listing any fitted parameters, background mathematical axioms, or prior results that the central claim depends on.

invented entities (1)
  • Observed Transition Primitives no independent evidence
    purpose: Decompose each transition into a sparse set to serve as interface for abstracting motion into action-like latents
    New intermediate representation introduced to handle ambiguity in multi-object scenes

pith-pipeline@v0.9.1-grok · 5715 in / 1208 out tokens · 39652 ms · 2026-06-30T05:58:22.371015+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

74 extracted references · 11 canonical work pages · 3 internal anchors

  1. [1]

    Proceedings of the 42nd International Conference on Machine Learning , year =

    Latent Action Learning Requires Supervision in the Presence of Distractors , author =. Proceedings of the 42nd International Conference on Machine Learning , year =

  2. [2]

    2021 , eprint=

    The Distracting Control Suite -- A Challenging Benchmark for Reinforcement Learning from Pixels , author=. 2021 , eprint=

    2021

  3. [3]

    ICLR 2026 the 2nd Workshop on World Models: Understanding, Modelling and Scaling , year=

    Hierarchical Latent Action Model , author=. ICLR 2026 the 2nd Workshop on World Models: Understanding, Modelling and Scaling , year=

    2026

  4. [4]

    2026 , eprint=

    You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences , author=. 2026 , eprint=

    2026

  5. [5]

    2022 , url=

    A Path Towards Autonomous Machine Intelligence Version 0.9.2, 2022-06-27 , author=. 2022 , url=

    2022

  6. [6]

    International Conference on Learning Representations , year=

    CLEVRER: Collision Events for Video Representation and Reasoning , author=. International Conference on Learning Representations , year=

  7. [7]

    The Fourteenth International Conference on Learning Representations , year=

    Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling , author=. The Fourteenth International Conference on Learning Representations , year=

  8. [8]

    Recurrent World Models Facilitate Policy Evolution , url =

    Ha, David and Schmidhuber, J\". Recurrent World Models Facilitate Policy Evolution , url =. Advances in Neural Information Processing Systems , editor =

  9. [9]

    2026 , eprint=

    Causal-JEPA: Learning World Models through Object-Level Latent Interventions , author=. 2026 , eprint=

    2026

  10. [10]

    The Eleventh International Conference on Learning Representations , year=

    SlotFormer: Unsupervised Visual Dynamics Simulation with Object-Centric Models , author=. The Eleventh International Conference on Learning Representations , year=

  11. [11]

    Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

    Locatello, Francesco and Weissenborn, Dirk and Unterthiner, Thomas and Mahendran, Aravindh and Heigold, Georg and Uszkoreit, Jakob and Dosovitskiy, Alexey and Kipf, Thomas , title =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =. 2020 , isbn =

    2020

  12. [12]

    The Twelfth International Conference on Learning Representations , year=

    Learning to Act without Actions , author=. The Twelfth International Conference on Learning Representations , year=

  13. [13]

    The Thirteenth International Conference on Learning Representations , year=

    Latent Action Pretraining from Videos , author=. The Thirteenth International Conference on Learning Representations , year=

  14. [14]

    2022 , eprint=

    Unsupervised Image Representation Learning with Deep Latent Particles , author=. 2022 , eprint=

    2022

  15. [15]

    Proceedings of the IEEE , volume=

    Toward causal representation learning , author=. Proceedings of the IEEE , volume=. 2021 , publisher=

    2021

  16. [17]

    2025 , url=

    Gaoyue Zhou and Hengkai Pan and Yann LeCun and Lerrel Pinto , booktitle=. 2025 , url=

    2025

  17. [18]

    Transactions on Machine Learning Research , issn=

    Maxime Oquab and Timoth. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  18. [19]

    2025 , eprint=

    Autoregressive Video Autoencoder with Decoupled Temporal and Spatial Context , author=. 2025 , eprint=

    2025

  19. [20]

    2025 , eprint=

    DeCo-VAE: Learning Compact Latents for Video Reconstruction via Decoupled Representation , author=. 2025 , eprint=

    2025

  20. [21]

    2025 , eprint=

    Hi-VAE: Efficient Video Autoencoding with Global and Detailed Motion , author=. 2025 , eprint=

    2025

  21. [22]

    arXiv preprint arXiv:2310.01040 , year =

    Segmenting the Motion Components of a Video: A Long-Term Unsupervised Model , author =. arXiv preprint arXiv:2310.01040 , year =

    work page arXiv

  22. [23]

    European Conference on Computer Vision , pages=

    Un-EVIMO: Unsupervised event-based independent motion segmentation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  23. [24]

    Choudhury, Subhabrata and Karazija, Laurynas and Laina, Iro and Vedaldi, Andrea and Rupprecht, Christian , booktitle =

  24. [25]

    The Fourteenth International Conference on Learning Representations , year=

    Articulation in Motion: Prior-free Part Mobility Analysis for Articulated Objects By Dynamic-Static Disentanglement , author=. The Fourteenth International Conference on Learning Representations , year=

  25. [26]

    NeurIPS 2025 Workshop on Embodied World Models for Decision Making , year=

    FLAM: Scaling Latent Action World Models with Factorization , author=. NeurIPS 2025 Workshop on Embodied World Models for Decision Making , year=

    2025

  26. [27]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Object-centric latent action learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  27. [28]

    2026 , eprint=

    Segment to Focus: Guiding Latent Action Models in the Presence of Distractors , author=. 2026 , eprint=

    2026

  28. [29]

    2026 , eprint=

    Learning Additively Compositional Latent Actions for Embodied AI , author=. 2026 , eprint=

    2026

  29. [30]

    Advances in neural information processing systems , volume=

    Neural discrete representation learning , author=. Advances in neural information processing systems , volume=

  30. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Generating holistic 3d human motion from speech , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  31. [32]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    X-dancer: Expressive music to human dance video generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  32. [33]

    European Conference on Computer Vision , pages=

    Nymeria: A massive collection of multimodal egocentric daily motion in the wild , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  33. [34]

    Transactions on Machine Learning Research , issn=

    Image Compression with Product Quantized Masked Image Modeling , author=. Transactions on Machine Learning Research , issn=. 2023 , url=

    2023

  34. [35]

    Proceedings of the 32nd International Conference on Machine Learning , year =

    Unsupervised Learning of Video Representations Using LSTMs , author =. Proceedings of the 32nd International Conference on Machine Learning , year =

  35. [36]

    2026 , eprint=

    Learning Latent Action World Models In The Wild , author=. 2026 , eprint=

    2026

  36. [37]

    Nature , volume=

    Mastering diverse control tasks through world models , author=. Nature , volume=. 2025 , publisher=

    2025

  37. [38]

    2024 , url=

    Nicklas Hansen and Hao Su and Xiaolong Wang , booktitle=. 2024 , url=

    2024

  38. [39]

    Conference on Robot Learning (CoRL) , year=

    What Matters in Learning from Offline Human Demonstrations for Robot Manipulation , author=. Conference on Robot Learning (CoRL) , year=

  39. [40]

    Proceedings of the Conference on Robot Learning , pages =

    Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning , author =. Proceedings of the Conference on Robot Learning , pages =. 2020 , editor =

    2020

  40. [41]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

  41. [42]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Film: Visual reasoning with a general conditioning layer , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  42. [43]

    Journal of Machine Learning Research , year =

    Laurens van der Maaten and Geoffrey Hinton , title =. Journal of Machine Learning Research , year =

  43. [44]

    Horn and Brian G

    Berthold K.P. Horn and Brian G. Schunck , abstract =. Determining optical flow , journal =. 1981 , issn =. doi:https://doi.org/10.1016/0004-3702(81)90024-2 , url =

    work page doi:10.1016/0004-3702(81)90024-2 1981

  44. [45]

    and Kanade, Takeo , title =

    Lucas, Bruce D. and Kanade, Takeo , title =. Proceedings of the 7th International Joint Conference on Artificial Intelligence - Volume 2 , pages =. 1981 , publisher =

    1981

  45. [46]

    1990 , url=

    An Isotropic 3×3 image gradient operator , author=. 1990 , url=

    1990

  46. [47]

    ArXiv , year=

    LAOF: Robust Latent Action Learning with Optical Flow Constraints , author=. ArXiv , year=

  47. [48]

    ArXiv , year=

    Learning Latent Action World Models In The Wild , author=. ArXiv , year=

  48. [49]

    arXiv preprint arXiv:2101.02722 , year =

    The Distracting Control Suite -- A Challenging Benchmark for Reinforcement Learning from Pixels , author =. arXiv preprint arXiv:2101.02722 , year =

    work page arXiv

  49. [50]

    , booktitle =

    Yi, Kexin and Gan, Chuang and Li, Yunzhu and Kohli, Pushmeet and Wu, Jiajun and Torralba, Antonio and Tenenbaum, Joshua B. , booktitle =. 2020 , url =

    2020

  50. [51]

    International Conference on Learning Representations , year =

    Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling , author =. International Conference on Learning Representations , year =

  51. [52]

    Advances in Neural Information Processing Systems , volume =

    Recurrent World Models Facilitate Policy Evolution , author =. Advances in Neural Information Processing Systems , volume =. 2018 , url =

    2018

  52. [53]

    Nam, Heejeong and Le Lidec, Quentin and Maes, Lucas and LeCun, Yann and Balestriero, Randall , journal =. Causal-. 2026 , url =

    2026

  53. [54]

    International Conference on Learning Representations , year =

    SlotFormer: Unsupervised Visual Dynamics Simulation with Object-Centric Models , author =. International Conference on Learning Representations , year =

  54. [55]

    Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

    Locatello, Francesco and Weissenborn, Dirk and Unterthiner, Thomas and Mahendran, Aravindh and Heigold, Georg and Uszkoreit, Jakob and Dosovitskiy, Alexey and Kipf, Thomas , title =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =. 2020 , publisher =

    2020

  55. [56]

    The Twelfth International Conference on Learning Representations , year =

    Learning to Act without Actions , author =. The Twelfth International Conference on Learning Representations , year =

  56. [57]

    Latent Action Pretraining from Videos

    Latent Action Pretraining from Videos , author =. arXiv preprint arXiv:2410.11758 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  57. [58]

    Proceedings of the 39th International Conference on Machine Learning , year =

    Unsupervised Image Representation Learning with Deep Latent Particles , author =. Proceedings of the 39th International Conference on Machine Learning , year =

  58. [59]

    Proceedings of the IEEE , volume =

    Toward causal representation learning , author =. Proceedings of the IEEE , volume =. 2021 , publisher =

    2021

  59. [60]

    Nonlinear independent component analysis: Existence and uniqueness results , journal =

    Hyv. Nonlinear independent component analysis: Existence and uniqueness results , journal =. 1999 , issn =. doi:https://doi.org/10.1016/S0893-6080(98)00140-3 , url =

    work page doi:10.1016/s0893-6080(98)00140-3 1999

  60. [61]

    arXiv preprint arXiv:2512.11293 , year =

    Autoregressive Video Autoencoder with Decoupled Temporal and Spatial Context , author =. arXiv preprint arXiv:2512.11293 , year =

    work page arXiv

  61. [62]

    2025 , url =

    Yin, Xiangchen and Yuan, Jiahui and Hu, Zhangchi and Sun, Wenzhang and Chen, Jie and Qiao, Xiaozhen and Li, Hao and Sun, Xiaoyan , journal =. 2025 , url =

    2025

  62. [63]

    Liu, Huaize and Sun, Wenzhang and Zhang, Qiyuan and Di, Donglin and Gong, Biao and Li, Hao and Wei, Chen and Zou, Changqing , journal =. Hi-. 2025 , url =

    2025

  63. [64]

    2024 , note =

    Wang, Ziyun and Guo, Jinyuan and Daniilidis, Kostas , booktitle =. 2024 , note =

    2024

  64. [65]

    British Machine Vision Conference , year =

    Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion , author =. British Machine Vision Conference , year =

  65. [66]

    NeurIPS 2025 Workshop on Embodied World Models for Decision Making , year =

    FLAM: Scaling Latent Action World Models with Factorization , author =. NeurIPS 2025 Workshop on Embodied World Models for Decision Making , year =

    2025

  66. [67]

    International Conference on Learning Representations , year =

    Object-Centric Latent Action Learning , author =. International Conference on Learning Representations , year =

  67. [68]

    Segment to Focus: Guiding Latent Action Models in the Presence of Distractors

    Segment to Focus: Guiding Latent Action Models in the Presence of Distractors , author =. arXiv preprint arXiv:2602.02259 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  68. [69]

    Learning Additively Compositional Latent Actions for Embodied AI

    Learning Additively Compositional Latent Actions for Embodied AI , author =. arXiv preprint arXiv:2604.03340 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  69. [70]

    Advances in Neural Information Processing Systems , volume =

    Neural Discrete Representation Learning , author =. Advances in Neural Information Processing Systems , volume =. 2017 , url =

    2017

  70. [71]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Generating Holistic 3D Human Motion from Speech , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  71. [72]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year =

    X-Dancer: Expressive Music to Human Dance Video Generation , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , year =

  72. [73]

    arXiv preprint arXiv:2406.09905 , year =

    Nymeria: A Massive Collection of Multimodal Egocentric Daily Motion in the Wild , author =. arXiv preprint arXiv:2406.09905 , year =

    work page arXiv

  73. [74]

    arXiv preprint arXiv:2212.07372 , year =

    Image Compression with Product Quantized Masked Image Modeling , author =. arXiv preprint arXiv:2212.07372 , year =

    work page arXiv

  74. [75]

    arXiv preprint arXiv:2601.05230 , year =

    Learning Latent Action World Models In The Wild , author =. arXiv preprint arXiv:2601.05230 , year =

    work page arXiv