Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

WorldPrediction: A Benchmark for High-level World Modeling and Long-horizon Procedural Planning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read WorldPrediction is a video-only, architecture-agnostic benchmark for high-level world modeling and procedural planning; on it, current frontier models reach 57% and 38% accuracy while humans are perfect.

desk verdict A useful new video benchmark for world modeling and planning, but the 'humans perfect' result is an artifact of the filtering protocol and Section 4.3 contradicts Table 2. read the letter →

arxiv 2506.04363 v1 pith:RT7WPGQR submitted 2025-06-04 cs.CV

Delong Chen , Willy Chung , Yejin Bang , Ziwei Ji , Pascale Fung This is my paper
classification cs.CV
keywords worldmodelingproceduralplanningbenchmarkvideounderstandingactionequivalentspartiallyobservablesemi-MDPlong-horizonmultiple-choiceevaluation
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

WorldPrediction is a video-only benchmark that tests whether AI systems can predict which high-level action turns one visual state into another (WorldPrediction-WM) and which ordered sequence of actions does so over three to ten steps (WorldPrediction-PP). The benchmark's central claim is that this can be evaluated fairly across architectures by presenting the model with pairs of before/after images and candidate action clips, with each candidate drawn from a different environment so that background continuity cannot give away the answer. On the human-filtered benchmark, the best current models reach 57.0% accuracy on WM and 38.1% on PP, against perfect human performance, which the paper takes as evidence that frontier models still lack robust causal understanding of abstract actions. If the benchmark is valid, it gives the field a common, architecture-agnostic yardstick for progress in high-level world modeling and long-horizon procedural planning from raw video.

What carries the argument

The load-bearing mechanism is 'action equivalents': for every high-level action category, the benchmark uses video clips of that action performed in visually different contexts, so the correct answer cannot be found by matching backgrounds, colors, or other task-irrelevant continuity cues. This is paired with a partially observable semi-MDP formulation (POSMDP): states and actions are latent, observations are images and videos, and evaluation is defined as an argmax over candidate actions or action sequences, i.e., which action category or ordered sequence is most likely to have produced the observed transition. The formulation supports both single-transition world modeling and multi-step procedural planning, where intermediate states are unobserved and must be inferred.

What would settle it

Take each WM sample and have independent annotators watch only the candidate action equivalents, then state what object or state change each clip produces; if a sizable fraction of 'equivalent' clips change different objects or states than the images require, those samples' gold answers are underdetermined and the reported accuracy cannot be a clean measure of world modeling.

Watch

Extended reading notes

Core claim

The paper claims that WorldPrediction is the first benchmark for high-level world modeling and procedural planning that evaluates semantic and temporal abstraction of actions purely from visual observations, without relying on text labels as model input. Its tasks are discriminative: given initial and final state images, a model must select the correct action clip (WM) or the correctly ordered action sequence (PP) among counterfactual distractors. To block shortcut learning, 'action equivalents'—clips sharing the same action label but recorded in different environments or viewpoints—replace the ground-truth action observations. The authors report that after human filtering, human annotators solve both tasks perfectly while state-of-the-art models score at most 57.0% on WM and 38.1% on PP, concluding that current systems have not mastered action-state causality at the semantic level.

Load-bearing premise

The benchmark's answers are only well-defined if clips carrying the same action label across datasets really produce the same causal transition from the initial to the final state; where labels are coarse or inconsistent, the correct action can be ambiguous.

Editorial extensions

If this is right

  • Current generative video models at roughly 26–30% on WM have not learned high-level action-state causality in pixel space, despite progress in lower-level video prediction.
  • Any new world model or planner can be evaluated on the same multiple-choice samples without retraining, because the benchmark is architecture-agnostic and takes only images and video as input.
  • Increasing model scale helps WorldPrediction-WM but not WorldPrediction-PP, so long-horizon procedural planning appears to need something other than larger perception and reasoning alone.
  • The label-free, video-only setup means performance comparisons no longer depend on human action annotations being fed to the model as input.

Reading between the lines

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

  • The paper's 'action equivalents' are built by matching textual labels across datasets, so the equivalence inherits whatever ambiguity those labels carry; two clips labeled the same can still depict different object-state changes, which would make some gold answers underdetermined.
  • A direct stress test would be to re-match equivalents by the state change they produce rather than by label and rerun the human filtering; a large drop in retained samples would show that label ambiguity, not world modeling, is part of the measured gap.
  • Because the WM task can be solved by tracking which objects change between the two images, an extension that varies the distractor actions while keeping object changes identical would separate object-state perception from true action understanding.
  • The PP task accepts multiple plausible plans as correct, so a natural extension is to score models by agreement with the human-preferred ordering rather than by exact match to one ground-truth sequence.
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

4 major / 5 minor

Summary. The paper introduces WorldPrediction, a video-based multiple-choice benchmark for high-level world modeling (WM) and long-horizon procedural planning (PP). It samples state-image pairs and action clips from COIN, CrossTask, EgoExo4D, EPIC-KITCHENS-100, and IKEA-ASM, represents actions via 'action equivalents' to block background-continuity shortcuts, filters samples by DINOv2 feature distances and human double-annotation, and evaluates VLMs, Socratic LLMs, video diffusion models, and OEPP planners. The headline claim is that frontier models reach at most 57.0% WM / 38.1% PP accuracy while humans solve both tasks perfectly.

Significance. If the benchmark is valid, it is a useful contribution: it is architecture-agnostic, purely visual, spans diverse real-world activities, includes long horizons up to 10 steps, and explicitly attempts to control for low-level visual shortcuts via action equivalents and observability filtering. The systematic comparison across model families is also valuable. However, the headline human-model gap is not currently an empirical result: human 'perfect' performance is enforced by the retention criterion in §4.3, and one of the cited model numbers contradicts the main table. These issues are fixable but must be addressed before the central claim can be accepted.

major comments (4)
  1. [§4.3 and Abstract/Conclusion] The statement that 'humans are able to solve both tasks perfectly' is circular. §4.3 says only samples where both annotators independently selected the correct answer were retained, so human accuracy on the released 825/570 samples is 100% by construction. The abstract and conclusion nevertheless report perfect human performance as evidence of a large human-model gap. This is load-bearing because the gap is the paper's headline result. Please reframe: report human performance on the original 1,500-sample pool (where inter-annotator agreement was 73% WM / 65% PP, Appendix A.1) or state explicitly that the filtering criterion selects samples that are human-solvable, and adjust the abstract and conclusion accordingly.
  2. [§4.3 vs Table 2 and Table 3] There is an internal inconsistency in the reported best model performance. §4.3 states 'the best current model performance, Claude-3.5 on WorldPrediction-WM, achieves only 45% accuracy,' but Table 2 lists Claude-3.5-Sonnet at 53.3% on WM and Qwen2.5-VL (72B) at 57.0% on WM. Similarly, Table 3 reports an overall PP accuracy of 37.4% for Llama-3.3 (70B) and 36.1% for Qwen2.5-VL (72B), while Table 2 shows 35.1% and 36.7% for the same models. Please correct these numbers and ensure the abstract's '57% / 38%' claim is consistent with the canonical tables.
  3. [§3.4 and §3.3, Action Equivalents] The validity of the benchmark rests on the assumption that clips sharing a textual action label across COIN, CrossTask, EPIC-KITCHENS-100, and IKEA-ASM are 'action equivalents' in the sense of inducing the same state transition. The appendix itself notes that EPIC-KITCHENS labels are fine-grained ('pick up') while COIN labels are macroscopic ('add', 'mix'), so identical labels may correspond to different environmental changes. If two equivalents produce different transitions, the intended answer is underdetermined even after human filtering, and the two-annotator filter may simply discard such ambiguous cases. Please provide evidence that same-label clips are causally equivalent (e.g., a small human study of state-change consistency per label, or a manual verification of the equivalent sets used in the released samples), or restrict equivalence to labels verified to induce the same transition.
  4. [§3.3, Observability Filtering] The DINOv2 distance thresholds (2.75 for WM, 10 for PP) are introduced without justification or sensitivity analysis. Because this filtering step removes a large fraction of the initially constructed samples and relies on a pretrained encoder's notion of semantic distance, the claim that the remaining samples are reliable and robust needs supporting evidence. Please report how the sample size and downstream model accuracies vary with the threshold, or otherwise justify the chosen values.
minor comments (5)
  1. [§4.2, Table 2 discussion] The sentence 'resulting in 25% and 77% unparsable responses in WM and PP, respectively' is ambiguous; please specify which model and task each percentage refers to.
  2. [Eq. (3), §3.2] The notation introduces \hat P as the correct sequence but then uses P* in the argmax; please define the true plan and the candidate plan consistently.
  3. [§3.1, POSMDP] The acronym POSMDP is introduced with references to POMDP and POMCP; please provide a canonical reference for the combined partially observable semi-MDP formulation or define it explicitly, since the framework is used as theoretical grounding.
  4. [Figure 6 caption] The caption 'Generated sample using CogVideoX-I2V' does not describe what is shown; please add a sentence explaining the generation condition and what is being visualized.
  5. [Table 1 caption and general text] The dataset name is inconsistent: 'EPIC-KITCHEN-100' appears in the Table 1 caption while 'EPIC-KITCHENS-100' is used elsewhere; please unify the spelling.

Circularity Check

1 steps flagged · score 6.0 of 10

Human 100% accuracy is a selection criterion, not a measured result: §4.3 keeps only samples both annotators answered correctly, then reports perfect human performance as evidence of a human-model gap.

  1. self definitional [Section 4.3 (Human Evaluation and Filtering); echoed in the Abstract and Section 5 (Conclusion)]
    "We adopted a conservative filtering criterion: only samples where both annotators independently provided the correct answer were retained. After filtering, we obtained 825 high-quality samples for WorldPrediction-WM and 570 samples for WorldPrediction-PP, ensuring that human performance was effectively perfect on the released benchmark."

    The released benchmark is defined as the subset of samples on which two annotators both chose the correct answer. Therefore 'human performance was effectively perfect' is true by construction: every sample that a human missed was deleted from the benchmark. The Abstract and Section 5 nevertheless present this filtered-in property as an empirical result ('humans are able to solve both tasks perfectly' / 'humans easily solving both tasks') that contrasts with frontier models' 57% and 38% accuracy. That specific human-model comparison is not an independent measurement; it is the definition of the curation rule. The model accuracies are unaffected and remain meaningful, but the headline perfect human score cannot serve as validation of benchmark solvability or as evidence of a human-model gap.

full rationale

The core model evaluation is not circular: the WM/PP tasks are constructed from external action-annotated datasets (COIN, CrossTask, EPIC-KITCHENS-100, EgoExo4D, IKEA-ASM), distractors are produced by shuffling or sampling alternative actions, and the reported model accuracies (57.0% WM, 38.1% PP) are measured against the resulting fixed benchmark without fitting to those results. There is no load-bearing self-citation chain; the POSMDP framing is standard and attributed to Kaelbling et al. and Sutton et al., and baselines come from external sources. The one clear circularity is the human performance headline. Section 4.3 explicitly adopts a filter that keeps only samples where both annotators independently selected the correct answer, then states this 'ensur[es] that human performance was effectively perfect on the released benchmark.' The Abstract and Conclusion then report that humans solve both tasks perfectly as an empirical contrast to frontier models. That contrast is forced by construction: human accuracy on the released set is 100% by definition, not by measurement. The inter-annotator agreement of 73% (WM) and 65% (PP) on the original 1500-sample pool suggests unfiltered human accuracy would be materially below perfect, but this number is not reported, so the gap is presented as larger than the data support. This makes the human-model gap partially circular even though the model scores themselves remain independently meaningful. A separate concern, that 'action equivalents' defined by shared textual labels may not be causally equivalent across datasets, is a validity and ambiguity risk rather than a circularity, so it is not scored here.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The benchmark depends on four unverified premises: semantic equivalence of same-label action clips across datasets, sufficiency of static state images, validity of DINOv2 distance as an observability filter, and the applicability of the POSMDP abstraction to real videos. Two thresholds are hand-set free parameters. No new physical entities are introduced.

free parameters (2)
  • DINOv2 distance thresholds = 2.75 (WM), 10 (PP)
    Used in observability filtering to discard sample pairs with large feature distances; values are stated without derivation and they change the composition of the benchmark.
  • Minimum EgoExo4D action segment length = 5 seconds
    Segments shorter than 5 seconds are discarded when building action equivalents; chosen by hand.
assumptions (4)
  • domain assumption High-level action categories in source datasets are semantically consistent across videos, so an action equivalent with the same label causes the same state transition.
    Action Equivalent Retrieval in Section 3.4 treats same textual label across COIN, CrossTask, EPIC-KITCHENS and IKEA as interchangeable; if labels are coarse or inconsistent, candidates may not be equivalent.
  • domain assumption Static initial and final images contain enough task-relevant visual evidence to determine which action caused the transition.
    Sample filtering in Section 3.3 keeps only pairs with small DINOv2 distance and visible interaction, assuming partial observability is controlled; ambiguous cases are removed by humans.
  • ad hoc to paper DINOv2 feature distance is a valid proxy for observability or noise in state pairs.
    Section 3.4 uses fixed thresholds to exclude drastic changes, but no validation is provided that this proxy selects causal transitions.
  • domain assumption The POSMDP formalism's latent state and action distinction applies to real instructional videos.
    Section 3.1 grounds the design in POSMDP; the benchmark treats video clips as observations of options and hidden transitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WorldPrediction: A Benchmark for High-level World Modeling and Long-horizon Procedural Planning." pith.science (2026). https://pith.science/paper/RT7WPGQR

@misc{pith2026250604363,
  author       = {Pith},
  title        = {Pith review of: WorldPrediction: A Benchmark for High-level World Modeling and Long-horizon Procedural Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RT7WPGQR}},
  note         = {Machine review of arXiv:2506.04363}
}
read the original abstract

Humans are known to have an internal "world model" that enables us to carry out action planning based on world states. AI agents need to have such a world model for action planning as well. It is not clear how current AI models, especially generative models, are able to learn such world models and carry out procedural planning in diverse environments. We introduce WorldPrediction, a video-based benchmark for evaluating world modeling and procedural planning capabilities of different AI models. In contrast to prior benchmarks that focus primarily on low-level world modeling and robotic motion planning, WorldPrediction is the first benchmark that emphasizes actions with temporal and semantic abstraction. Given initial and final world states, the task is to distinguish the proper action (WorldPrediction-WM) or the properly ordered sequence of actions (WorldPrediction-PP) from a set of counterfactual distractors. This discriminative task setup enable us to evaluate different types of world models and planners and realize a thorough comparison across different hypothesis. The benchmark represents states and actions using visual observations. In order to prevent models from exploiting low-level continuity cues in background scenes, we provide "action equivalents" - identical actions observed in different contexts - as candidates for selection. This benchmark is grounded in a formal framework of partially observable semi-MDP, ensuring better reliability and robustness of the evaluation. We conduct extensive human filtering and validation on our benchmark and show that current frontier models barely achieve 57% accuracy on WorldPrediction-WM and 38% on WorldPrediction-PP whereas humans are able to solve both tasks perfectly.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SWITCH: Benchmarking Modeling and Handling of Tangible Interfaces in Long-horizon Embodied Scenarios

    cs.CV 2025-11 conditional novelty 6.0 of 10

    SWITCH introduces a 193-video benchmark of tangible control-interface interactions and shows that frontier LMMMs struggle with fine-grained grounding and outcome verification.

  2. A Definition and Roadmap for World Models

    cs.AI 2026-07 conditional novelty 5.0 of 10

    A perspective article defining world models as finite-resource compression of physical state transitions and outlining a roadmap toward physical AGI via unified representations and interactive simulators.

  3. Embodied AI Agents: Modeling the World

    cs.AI 2025-06 conditional novelty 4.0 of 10

    Embodied AI agents should be built around physical world models plus a mental world model of the user, with virtual, wearable, and robotic agents sharing this core.

Reference graph

Works this paper leans on

22 extracted references · 7 canonical work pages · cited by 3 Pith papers

  1. [1]

    Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning

    Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning. InICLR 2024 Workshop on Large Language Model (LLM) Agents,

    work page 2024

  2. [6]

    World models.arXiv preprint arXiv:1803.10122,

    David Ha and Jürgen Schmidhuber. World models.arXiv preprint arXiv:1803.10122,

    arXiv

  3. [8]

    Mmworld: Towards multi-discipline multi-faceted world model evaluation in videos

    Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multi-discipline multi-faceted world model evaluation in videos. In Workshop on Video-Language Models@ NeurIPS 2024,

    work page 2024

  4. [9]

    Gaia-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080,

    Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080,

    arXiv

  5. [10]

    Text2world: Benchmarking large language models for symbolic world model generation

    Mengkang Hu, Tianxing Chen, Yude Zou, Yuheng Lei, Qiguang Chen, Ming Li, Qiwei Liang, Yao Mu, Hongyuan Zhang, Wenqi Shao, et al. Text2world: Benchmarking large language models for symbolic world model generation. InICLR 2025 Workshop on World Models: Understanding, Modelling and Scaling,

    work page 2025

  6. [12]

    Worldmodelbench: Judging video generation models as world models.arXiv preprint arXiv:2502.20694,

    Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu Yin, Joseph E Gonzalez, et al. Worldmodelbench: Judging video generation models as world models.arXiv preprint arXiv:2502.20694,

    arXiv

  7. [13]

    A language-first approach for procedure planning

    Jiateng Liu, Sha Li, Zhenhailong Wang, Manling Li, and Heng Ji. A language-first approach for procedure planning. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors,Findings of the Association for Computational Linguistics: ACL 2023, pages 1941–1954, Toronto, Canada, July

    work page 2023

  8. [14]

    doi: 10.18653/v1/2023.findings-acl.122.https://aclanthology.org/2023.findings-acl.122/

    Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-acl.122.https://aclanthology.org/2023.findings-acl.122/. Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time.IEEE Robotics and Automation Letters,

    work page doi:10.18653/v1/2023.findings-acl.122.https://aclanthology.org/2023.findings-acl.122/ 2023

Show all 22 references
  1. [15]

    Cosmos world foundation model platform for physical ai, 2025.https://arxiv.org/abs/2501.03575

    NVIDIA, :, Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, Daniel Dworakowski, Jiaojiao Fan, Michele Fenzi, Francesco Ferroni, Sanja Fidler, Dieter Fox, Songwei Ge, Yunhao Ge, Jinwei ...

    2025 arXiv

  2. [17]

    Open-event procedure planning in instructional videos.arXiv preprint arXiv:2407.05119,

    Yilu Wu, Hanlin Wang, Jing Wang, and Limin Wang. Open-event procedure planning in instructional videos.arXiv preprint arXiv:2407.05119,

    arXiv

  3. [18]

    VideoCLIP: Contrastive pre-training for zero-shot video-text understanding

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. InProceedings of the 2021 Conference on Empirical Methods in Natural La...

    2021

  4. [19]

    Qwen2 technical report.CoRR, 2024a

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report.CoRR, 2024a. Chang Yang, Xinrun Wang, Junzhe Jiang, Qinggang Zhang, and Xiao Huang. Evaluating world models with llm for decis...

    2023

  5. [20]

    Socratic models: Composing zero-shot multimodal reasoning with language.arXiv preprint arXiv:2204.00598,

    Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language.arXiv preprint arXiv:2204.00598,

    arXiv

  6. [21]

    I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models.arXiv preprint arXiv:2311.04145,

    Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models.arXiv preprint arXiv:2311.04145,

    arXiv

  7. [22]

    Dino-wm: World models on pre-trained visual features enable zero-shot planning.arXiv preprint arXiv:2411.04983, 2024a

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. Dino-wm: World models on pre-trained visual features enable zero-shot planning.arXiv preprint arXiv:2411.04983, 2024a. Siyuan Zhou, Yilun Du, Jiaben Chen, Yandong Li, Dit-Yan Yeung, and Chuang Gan. Robodreamer: learning c...

    arXiv

  8. [2017]

    Event-guided procedure planning from instructional videos with text supervision

    An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13565–13575, 2023a. Hanlin Wang, Yilu Wu, Shen...

    2023

  9. [2019]

    Mastering diverse domains through world models

    12 Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104,

    arXiv

  10. [2021]

    A path towards autonomous machine intelligence version 0.9

    Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27.Open Review, 62(1):1–62,

    2022

  11. [2022]

    Worldscore: A unified evaluation benchmark for world generation.arXiv preprint arXiv:2504.00983,

    Haoyi Duan, Hong-Xing Yu, Sirui Chen, Li Fei-Fei, and Jiajun Wu. Worldscore: A unified evaluation benchmark for world generation.arXiv preprint arXiv:2504.00983,

  12. [2023]

    Navigation world models.arXiv preprint arXiv:2412.03572,

    Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models.arXiv preprint arXiv:2412.03572,

    arXiv

  13. [2024]

    Intuitive physics understanding emerges from self-supervised pretraining on natural videos.arXiv preprint arXiv:2502.11831,

    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.arXiv preprint arXiv:2502.11831,

    arXiv

  14. [2025]

    Learning and leveraging world models in visual representation learning.arXiv preprint arXiv:2403.00504,

    Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, and Yann LeCun. Learning and leveraging world models in visual representation learning.arXiv preprint arXiv:2403.00504,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.