pith. machine review for the scientific record. sign in

arxiv: 2605.06388 · v1 · submitted 2026-05-07 · 💻 cs.CV · cs.LG· cs.RO

Recognition: unknown

Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models

Nilaksh , Saurav Jha , Artem Zholus , Sarath Chandar
Authors on Pith no claims yet

Pith reviewed 2026-05-08 13:12 UTC · model grok-4.3

classification 💻 cs.CV cs.LGcs.RO
keywords robotic world modelslatent diffusion modelssemantic encodersreconstruction encodersaction-conditioned videopolicy learninglatent space evaluation
0
0 comments X

The pith

Semantic latent spaces outperform reconstruction ones for robotic world models even with weaker pixel fidelity.

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

The paper compares six different encoders to build latent spaces for action-conditioned video diffusion models that serve as robotic world simulators. It trains these models on the same BridgeV2 robot data and measures results along visual image quality, ability to plan actions, and quality of the learned representations. Reconstruction encoders such as VAEs score well on pixel accuracy but fall short on planning and policy tasks, while semantic encoders like V-JEPA 2.1, Web-DINO, and SigLIP 2 produce stronger policy learning and better latent properties at every scale tested. This leads to the conclusion that semantic understanding in the latent space matters more than perfect visual reconstruction when the goal is useful robot control.

Core claim

Semantic encoders such as V-JEPA 2.1 (strongest on policy), Web-DINO, and SigLIP 2 generally outperform reconstruction encoders like VAE and Cosmos on planning, downstream policy performance, and latent quality across model scales, even though the reconstruction models achieve higher visual fidelity scores. High-dimensional semantic spaces support effective world model training with or without compression. Visual fidelity alone is therefore insufficient for selecting world models intended for robotics.

What carries the argument

The three evaluation axes of visual fidelity, planning and downstream policy performance, and latent representation quality applied to action-conditioned latent diffusion models built from different pretrained encoders.

If this is right

  • Visual fidelity metrics cannot be trusted by themselves to pick the best world model for robotics.
  • Semantic encoders deliver better planning and policy outcomes at every tested model scale.
  • World models can be trained successfully in high-dimensional semantic spaces without dimension compression.
  • Pretrained semantic encoders supply stronger starting points than reconstruction autoencoders for diffusion-based robotic simulators.

Where Pith is reading between the lines

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

  • Future robotic simulators could prioritize semantic pretraining objectives over pure pixel reconstruction when designing latent spaces.
  • The advantage of semantic spaces may appear in other control tasks where relations between objects matter more than exact visual appearance.
  • Direct tests on physical robots would show whether the observed policy gains transfer from simulation to hardware.

Load-bearing premise

That performance differences seen on the BridgeV2 dataset under one fixed training protocol will determine the best latent space for robotic world models in general.

What would settle it

A controlled experiment on a different robot dataset or with varied training protocols in which reconstruction encoders produce higher policy success rates than semantic encoders.

Figures

Figures reproduced from arXiv: 2605.06388 by Artem Zholus, Nilaksh, Sarath Chandar, Saurav Jha.

Figure 1
Figure 1. Figure 1: Which latent space makes a better robotic world model? For a latent diffusion model, we fix the Diffusion Transformer (DiT) transition model, action conditioning, and training data. We vary only the encoder fϕ defined latent interface: encoder, optional compression adapter, and the associated decoder path. This isolates how reconstruction-aligned and semantic representations affect action-faithful dynamics… view at source ↗
Figure 2
Figure 2. Figure 2: Action trajectories induced by encoder spaces: episode rollouts projected onto the top-2 canonical-correlation directions between IDM features and ground-truth actions. (ρ1, ρ2) are the leading canonical correla￾tions, η summarizes the aggregate action alignment. Colored curves are episodes. Rather than predicting future frames directly in pixel space, latent world models learn predictive dynamics in a rep… view at source ↗
Figure 3
Figure 3. Figure 3: Latent space effect overview: each point is a DiT-S world model trained by varying only the en￾coder and the associated decoder path. (a) Upper-right is favorable. Latent space metrics show that semantic encoders improve action recoverability, task-success separability, and action planning error (CEM) relative to reconstruction-aligned encoders. (b) Lower-right is favorable. Visual utility metrics show tha… view at source ↗
Figure 4
Figure 4. Figure 4: SSIM gap over steps. Semantic latent spaces remain visually competitive view at source ↗
Figure 5
Figure 5. Figure 5: Scaling camera views (left) and DiT sizes (right). Multi-view training improves action recovery but can hurt video qual￾ity under limited data. We take the trained DiT-S models and finetune them for 20 epochs on the BridgeV2 episodes that contain three camera views view at source ↗
Figure 6
Figure 6. Figure 6: Adapter ablation results. The main failure modes differ: reconstruction latents hal￾lucinate task semantics, while semantic latents miss geom￾etry and contact. Our qualitative rollouts in Appx. fig. 13 show that all encoder families share a common failure mode where static scene elements are faithfully preserved while manipulation-relevant details hallucinate. Beyond this univer￾sal pattern, encoder famili… view at source ↗
Figure 7
Figure 7. Figure 7: Open-VLA success rate comparison on two random episodes: four frames are sampled at even intervals. ✓ and ✗ show trajectories marked as success and failure by InternVL 3.5 VLM. Vid2World [25] causalizes video diffusion with action guidance; Ctrl-World [19] studies multi-view, long￾horizon, policy-in-the-loop manipulation. A third line moves prediction and planning into semantic feature space: DINO-WM [80],… view at source ↗
Figure 8
Figure 8. Figure 8: The Cohen’s kappa for inter-VLM rater agreement. Given the higher agreement between InternVL 3.5 and Qwen 3.6, we choose these as our VLM judges for policy-in-the-loop task success experiments. D Additional Results 27 view at source ↗
Figure 9
Figure 9. Figure 9: SSIM gap, LPIPS gap, and PCK coverage over 45 rollout steps. While all encoders show a strictly increasing SSIM/LPIPS gap over the full rollout due to compounding errors (each autoregressive step feeds back slightly corrupted predictions as context), semantic latent spaces from SigLIP2, V-JEPA 2.1 and Web-DINO remain particularly competitive when forced to extrapolate beyond the 10-frame horizon length see… view at source ↗
Figure 10
Figure 10. Figure 10: Action trajectories induced by encoder spaces: episode rollouts projected onto the top-2 canonical-correlation directions between IDM features and ground-truth actions. (ρ1, ρ2) are the leading canonical correlations, η summarizes the aggregate action alignment. Colored curves are episodes. D.5 Multi-view transfer learning view at source ↗
Figure 11
Figure 11. Figure 11: Open-VLA success rate comparison on two random episodes: four frames are sampled at even intervals. ✓ and ✗ show trajectories marked as success and failure by InternVL 3.5 VLM. 34 view at source ↗
Figure 12
Figure 12. Figure 12: Pixel rollout comparison across models on diverse episodes: the first frame is fed as context and the rest 3 frames are sampled at even intervals from the generated world model rollout. 35 view at source ↗
Figure 13
Figure 13. Figure 13: Hallucinated pixel rollout comparison across models on diverse episodes: the first frame is fed as context and the rest 3 frames are sampled at even intervals from the generated world model rollout. Top: flipping the pot consistently causes distortions for all models; Middle: turning the book pages causes the models to only partially follow the motion with the book/page appearances becoming smeared and in… view at source ↗
Figure 14
Figure 14. Figure 14: OOD Distractor comparison showing failure episodes per model: OOD objects break task￾object binding and action-conditioned state tracking across all models. ✓ and ✗ show trajectories marked as success and failure by InternVL 3.5. In their respective trajectories: Cosmos generates less stable towel/object state; VAE fails at task-relevant placement of the silver pot; V-JEPA 2.1 loses stable binding between… view at source ↗
Figure 15
Figure 15. Figure 15: OOD Distractor comparison for the same episode: OOD distractor competes with the target objects and exposes whether a model can keep the target objects bound to the instruction. ✓ and ✗ show trajectories marked as success and failure by InternVL 3.5. Here, irrespective of the task success, the added object visually changes the predicted interaction for all models: the robot/can motion becomes less task￾di… view at source ↗
Figure 16
Figure 16. Figure 16: OOD Instruction comparison showing failure cases per model: for each model, the same initial context is rolled out with the original instruction, which succeeds, and then with a OOD instruction, which fails. ✓ and ✗ show trajectories marked as success and failure by InternVL 3.5. Model-specific OOD instruction trajectories show: Cosmos rollout still moves the blue scrubber around the stove, but does not r… view at source ↗
Figure 17
Figure 17. Figure 17: OOD Instruction comparison for the same episode: most models exhibit a common hallucina￾tion where the original object dynamics or defaults to a familiar action pattern instead of updating the final state to match the new instruction. ✓ and ✗ show trajectories marked as success and failure by InternVL 3.5. Both Cosmos and VAE maintain the cloth in a partially folded/creased state instead of flattening it.… view at source ↗
read the original abstract

World model-based policy evaluation is a practical proxy for testing real-world robot control by rolling out candidate actions in action-conditioned video diffusion models. As these models increasingly adopt latent diffusion modeling (LDM), choosing the right latent space becomes critical. While the status quo uses autoencoding latent spaces like VAEs that are primarily trained for pixel reconstruction, recent work suggests benefits from pretrained encoders with representation-aligned semantic latent spaces. We systematically evaluate these latent spaces for action-conditioned LDM by comparing six reconstruction and semantic encoders to train world model variants under a fixed protocol on BridgeV2 dataset, and show effective world model training in high-dimensional representation spaces with and without dimension compression. We then propose three axes to assess robotic world model performance: visual fidelity, planning and downstream policy performance, and latent representation quality. Our results show visual fidelity alone is insufficient for world model selection. While reconstruction encoders like VAE and Cosmos achieve strong pixel-level scores, semantic encoders such as V-JEPA 2.1 (strongest overall on policy), Web-DINO, and SigLIP 2 generally excel across the other two axes at all model scales. Our study advocates semantic latent space as stronger foundation for policy-relevant robotics diffusion world models.

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 systematically compares six off-the-shelf encoders (reconstruction-based: VAE, Cosmos; semantic: V-JEPA 2.1, Web-DINO, SigLIP 2 and others) for action-conditioned latent diffusion models trained as robotic world models on the BridgeV2 dataset under a fixed protocol. It evaluates them on three axes (visual fidelity, planning/policy performance, latent representation quality) and concludes that visual fidelity alone is insufficient for selection, with semantic encoders generally excelling on the other axes at all model scales and with/without dimension compression, advocating semantic latent spaces as a stronger foundation for policy-relevant robotics diffusion world models.

Significance. If the empirical ranking holds under broader conditions, the work provides actionable guidance for latent-space selection in robotic world models, potentially improving policy transfer by prioritizing semantic alignment over pixel reconstruction.

major comments (2)
  1. [Abstract and Evaluation Protocol] The central claim that semantic encoders (V-JEPA 2.1 strongest on policy) provide a stronger foundation rests on results from a single dataset (BridgeV2) and one fixed training protocol. The abstract's assertion that advantages hold 'at all model scales' and 'with and without dimension compression' therefore requires explicit ablations or additional datasets to rule out protocol- or distribution-specific effects before supporting the general recommendation.
  2. [Results] No variance, standard deviations, or statistical significance tests are referenced for the cross-encoder rankings on the planning/policy axis. Without these, the reported superiority (e.g., V-JEPA 2.1 as strongest overall) cannot be assessed for robustness against run-to-run variability or hyperparameter sensitivity in high-dimensional semantic spaces.
minor comments (2)
  1. [Methods] Clarify the exact procedure and hyperparameters for dimension compression when applied to semantic encoders, and whether any implicit tuning favors higher-dimensional representations.
  2. [Results] Add a table or figure summarizing all three axes side-by-side for each encoder and scale to improve readability of the comparative claims.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below with clarifications, explanations of our experimental design, and indications of revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Evaluation Protocol] The central claim that semantic encoders (V-JEPA 2.1 strongest on policy) provide a stronger foundation rests on results from a single dataset (BridgeV2) and one fixed training protocol. The abstract's assertion that advantages hold 'at all model scales' and 'with and without dimension compression' therefore requires explicit ablations or additional datasets to rule out protocol- or distribution-specific effects before supporting the general recommendation.

    Authors: We agree that broader validation across datasets would strengthen the general recommendation. Our study deliberately uses a fixed protocol on BridgeV2 to enable controlled, apples-to-apples comparison of the six encoders. The claims regarding 'all model scales' and 'with and without dimension compression' are directly supported by the ablations we performed and reported within this setup. In the revised manuscript we will (1) revise the abstract to qualify the scope of the claims, (2) add an explicit limitations paragraph discussing single-dataset evaluation, and (3) include a forward-looking statement on the value of future multi-dataset studies. No new large-scale experiments are feasible within the revision timeline, but the existing scale and compression ablations already address part of the concern. revision: partial

  2. Referee: [Results] No variance, standard deviations, or statistical significance tests are referenced for the cross-encoder rankings on the planning/policy axis. Without these, the reported superiority (e.g., V-JEPA 2.1 as strongest overall) cannot be assessed for robustness against run-to-run variability or hyperparameter sensitivity in high-dimensional semantic spaces.

    Authors: We acknowledge that reporting variability would improve confidence in the rankings. Training each world-model variant is computationally expensive; therefore we report single-run results for the full set of encoders and scales. However, the superiority of semantic encoders is consistent across three independent evaluation axes (visual fidelity, planning/policy, representation quality) and across model scales, which mitigates the risk of isolated run-specific artifacts. In the revision we will add a dedicated limitations subsection on single-run evaluation, report standard deviations for the subset of smaller-scale models where we have multiple seeds, and note that hyperparameter sensitivity was controlled by using the same training protocol for all encoders. revision: partial

standing simulated objections not resolved
  • Full validation on additional robotic datasets beyond BridgeV2 to confirm the ranking is not distribution-specific

Circularity Check

0 steps flagged

No significant circularity: empirical encoder comparison

full rationale

The paper conducts a direct empirical comparison of six off-the-shelf encoders (VAE, Cosmos, V-JEPA 2.1, Web-DINO, SigLIP 2, and one other) by training action-conditioned LDM world models on BridgeV2 under a fixed protocol. It reports results on three evaluation axes without any mathematical derivations, parameter-fitting steps that are then relabeled as predictions, self-referential definitions, or load-bearing uniqueness theorems. All claims rest on observable performance differences across visual fidelity, policy/planning, and latent quality metrics, which are independent of the paper's own inputs by construction. Minor self-citations to prior encoder work are present but do not support the central ranking; the study remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on empirical results from training and evaluating world models on the BridgeV2 dataset; no free parameters, axioms beyond standard diffusion modeling assumptions, or invented entities are introduced.

pith-pipeline@v0.9.0 · 5526 in / 1107 out tokens · 53689 ms · 2026-05-08T13:12:06.999177+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

89 extracted references · 29 canonical work pages · 10 internal anchors

  1. [1]

    Cosmos World Foundation Model Platform for Physical AI

    Niket Agarwal, Arslan Ali, Maciej Bala, Y ogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopad- hyay, Y ongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025

    work page internal anchor Pith review arXiv 2025

  2. [2]

    Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation

    Jason Ansel, Edward Y ang, Horace He, Natalia Gimelshein, Animesh Jain, Michael V oznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM international conference on architectural support for programming la...

    2024

  3. [3]

    Self-supervised learning from images with a joint-embedding predic- tive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Y ann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predic- tive architecture. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 15619–15629, 2023

    2023

  4. [4]

    V -JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning, June 2025

    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, V asil Khalidov, Patrick Labatut, Fran- cisco Massa, Marc Szafraniec, Kapil Krishnakumar, Y ong Li,...

    2025

  5. [5]

    SemanticGen: Video Generation in Seman- tic Space, December 2025

    Jianhong Bai, Xiaoshi Wu, Xintao Wang, Xiao Fu, Y uanxing Zhang, Qinghe Wang, Xiaoyu Shi, Meng- han Xia, Zuozhu Liu, Haoji Hu, Pengfei Wan, and Kun Gai. SemanticGen: Video Generation in Seman- tic Space, December 2025

    2025

  6. [6]

    Qwen3-VL Technical Report

    Shuai Bai, Y uxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, 10 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng...

    work page internal anchor Pith review arXiv 2025

  7. [7]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Y uanzhi Zhu, Mingkun Y ang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Y e, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Y ang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical repor...

    work page internal anchor Pith review arXiv 2025

  8. [8]

    Back to the features: Dino as a foundation for video world models.arXiv preprint arXiv:2507.19468,

    Federico Baldassarre, Marc Szafraniec, Basile Terver, V asil Khalidov, Francisco Massa, Y ann LeCun, Patrick Labatut, Maximilian Seitzer, and Piotr Bojanowski. Back to the features: Dino as a foundation for video world models. arXiv preprint arXiv:2507.19468, 2025

    work page arXiv 2025

  9. [9]

    Model-predictive control via cross-entropy and gradient-based optimization

    Homanga Bharadhwaj, Kevin Xie, and Florian Shkurti. Model-predictive control via cross-entropy and gradient-based optimization. In Learning for Dynamics and Control , pages 277–286. PMLR, 2020

    2020

  10. [10]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Y ufei Guo, Li Jing, David Schnurr, Joe Tay- lor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai. com/research/video-generation-models-as-world-simulators, 3(1):3, 2024

    2024

  11. [11]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650–9660, 2021

    2021

  12. [12]

    Diffusion forcing: Next-token prediction meets full-sequence diffusion

    Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Informa- tion Processing Systems, 37:24081–24125, 2024

    2024

  13. [13]

    V ery deep {vae}s generalize autoregressive models and can outperform them on images

    Rewon Child. V ery deep {vae}s generalize autoregressive models and can outperform them on images. In International Conference on Learning Representations , 2021. URL https://openreview.net/ forum?id=RLRXCV6DbEJ

    2021

  14. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In In- ternational Conference on Learning Representations, 2021. URL htt...

    2021

  15. [15]

    Simulation tools for model-based robotics: Comparison of bullet, havok, mujoco, ode and physx

    Tom Erez, Y uval Tassa, and Emanuel Todorov. Simulation tools for model-based robotics: Comparison of bullet, havok, mujoco, ode and physx. In 2015 IEEE international conference on robotics and automation (ICRA), pages 4397–4404. IEEE, 2015

    2015

  16. [16]

    Scaling rectified flow transformers for high- resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Y am Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high- resolution image synthesis. In F orty-first international conference on machine learning, 2024

    2024

  17. [17]

    arXiv preprint arXiv:2601.04137 (2026)

    Chun-Kai Fan, Xiaowei Chi, Xiaozhu Ju, Hao Li, Y ong Bao, Y u-Kai Wang, Lizhang Chen, Zhiyuan Jiang, Kuangzhi Ge, Ying Li, et al. Wow, wo, val! a comprehensive embodied world model evaluation turing test. arXiv preprint arXiv:2601.04137, 2026

    work page arXiv 2026

  18. [18]

    Scaling language-free visual representation learning

    David Fan, Shengbang Tong, Jiachen Zhu, Koustuv Sinha, Zhuang Liu, Xinlei Chen, Michael Rabbat, Nicolas Ballas, Y ann LeCun, Amir Bar, et al. Scaling language-free visual representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 370–382, 2025. 11 Reconstruction or Semantics? What Makes a Latent Space Usefu...

    2025

  19. [19]

    Ctrl-world: A controllable generative world model for robot manipulation.arXiv preprint arXiv:2510.10125, 2025

    Y anjiang Guo, Lucy Xiaoyang Shi, Jianyu Chen, and Chelsea Finn. Ctrl-world: A controllable genera- tive world model for robot manipulation. arXiv preprint arXiv:2510.10125, 2025

    work page arXiv 2025

  20. [20]

    World Models

    David Ha and Jürgen Schmidhuber. World models. eprint arXiv: 1803.10122 , 2018

    work page internal anchor Pith review arXiv 2018

  21. [21]

    Learning latent dynamics for planning from pixels

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555–2565. PMLR, 2019

    2019

  22. [22]

    Momentum contrast for unsuper- vised visual representation learning

    Kaiming He, Haoqi Fan, Y uxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsuper- vised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  23. [23]

    Masked autoen- coders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Y anghao Li, Piotr Dollár, and Ross Girshick. Masked autoen- coders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022

    2022

  24. [24]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  25. [25]

    Huang, J

    Siqiao Huang, Jialong Wu, Qixing Zhou, Shangchen Miao, and Mingsheng Long. Vid2world: Crafting video diffusion models to interactive world models. arXiv preprint arXiv:2505.14357, 2025

    work page arXiv 2025

  26. [26]

    Cotracker: It is better to track together

    Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea V edaldi, and Christian Rupprecht. Cotracker: It is better to track together. In Proc. ECCV, 2024

    2024

  27. [27]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Y ang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision , pages 5148–5157, 2021

    2021

  28. [28]

    Openvla: An open-source vision- language-action model

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision- language-action model. In Conference on Robot Learning , pages 2679–2713. PMLR, 2025

    2025

  29. [29]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    work page internal anchor Pith review arXiv 2013

  30. [30]

    Imagenet classification with deep convolu- tional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolu- tional neural networks. Advances in neural information processing systems , 25, 2012

    2012

  31. [31]

    R-bench: Are your large multimodal model robust to real-world corruptions?, 2024

    Chunyi Li, Jianbo Zhang, Zicheng Zhang, Haoning Wu, Y uan Tian, Wei Sun, Guo Lu, Xiaohong Liu, Xiongkuo Min, Weisi Lin, and Guangtao Zhai. R-bench: Are your large multimodal model robust to real-world corruptions?, 2024

    2024

  32. [32]

    Gonzalez, Ion Stoica, Song Han, and Y ao Lu

    Dacheng Li, Y unhao Fang, Y ukang Chen, Shuo Y ang, Shiyi Cao, Justin Wong, Michael Luo, Xiao- long Wang, Hongxu Yin, Joseph E. Gonzalez, Ion Stoica, Song Han, and Y ao Lu. Worldmodel- bench: Judging video generation models as world models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , 2025....

    2025

  33. [33]

    Evaluating real-world robot manipulation policies in simula- tion

    Xuanlin Li, Kyle Hsu, Jiayuan Gu, Oier Mees, Karl Pertsch, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, et al. Evaluating real-world robot manipulation policies in simula- tion. In Conference on Robot Learning , pages 3705–3728. PMLR, 2025

    2025

  34. [34]

    Worldeval: World model as real-world robot policies evaluator.arXiv preprint arXiv:2505.19017, 2025

    Y axuan Li, Yichen Zhu, Junjie Wen, Chaomin Shen, and Yi Xu. Worldeval: World model as real-world robot policies evaluator. arXiv preprint arXiv:2505.19017, 2025. 12 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models

    work page arXiv 2025

  35. [35]

    Y aron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations , 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  36. [36]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  37. [37]

    VDT: General- purpose video diffusion transformers via mask modeling

    Haoyu Lu, Guoxing Y ang, Nanyi Fei, Y uqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. VDT: General- purpose video diffusion transformers via mask modeling. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=Un0rgm9f04

    2024

  38. [38]

    arXiv preprint arXiv:2603.14482 (2026)

    Lorenzo Mur-Labadia, Matthew Muckley, Amir Bar, Mido Assran, Koustuv Sinha, Mike Rabbat, Y ann LeCun, Nicolas Ballas, and Adrien Bardes. V -jepa 2.1: Unlocking dense features in video self- supervised learning. arXiv preprint arXiv:2603.14482, 2026

    work page arXiv 2026

  39. [39]

    Robocasa: Large-scale simulation of everyday tasks for generalist robots

    Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Y uke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems (RSS) , 2024

    2024

  40. [40]

    Robocasa365: A large-scale simulation framework for training and benchmarking generalist robots

    Soroush Nasiriany, Sepehr Nasiriany, Abhiram Maddukuri, and Y uke Zhu. Robocasa365: A large-scale simulation framework for training and benchmarking generalist robots. In International Conference on Learning Representations (ICLR), 2026

    2026

  41. [41]

    DINOv2: Learning Robust Visual Features without Supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, V asil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    work page internal anchor Pith review arXiv 2023

  42. [42]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 4195–4205, 2023

    2023

  43. [43]

    Sample-efficient cross-entropy method for real-time planning

    Cristina Pinneri, Shambhuraj Sawant, Sebastian Blaes, Jan Achterhold, Joerg Stueckler, Michal Rolinek, and Georg Martius. Sample-efficient cross-entropy method for real-time planning. In Conference on Robot Learning, pages 1049–1065. PMLR, 2021

    2021

  44. [44]

    Worldgym: World model as an environment for policy evaluation, 2025

    Julian Quevedo, Ansh Kumar Sharma, Yixiang Sun, V arad Suryavanshi, Percy Liang, and Sherry Y ang. Worldgym: World model as an environment for policy evaluation. arXiv preprint arXiv:2506.00613 , 2025

    work page arXiv 2025

  45. [45]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen.ai/ blog?id=qwen3.5

    2026

  46. [46]

    Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026

    Qwen Team. Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026. URL https: //qwen.ai/blog?id=qwen3.6-27b

    2026

  47. [47]

    Learning transferable visual models from nat- ural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. In International conference on machine learning , pages 8748–8763. PmLR, 2021

    2021

  48. [48]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684–10695, 2022

    2022

  49. [49]

    The cross-entropy method: a unified approach to combinato- rial optimization, Monte-Carlo simulation, and machine learning , volume 133

    Reuven Y Rubinstein and Dirk P Kroese. The cross-entropy method: a unified approach to combinato- rial optimization, Monte-Carlo simulation, and machine learning , volume 133. Springer, 2004

    2004

  50. [50]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard V encu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278–25294, 2022. 13 Reconstruction or Semantics? ...

    2022

  51. [51]

    arXiv preprint arXiv:2602.08971 (2026)

    Y u Shang, Zhuohang Li, Yiding Ma, Weikang Su, Xin Jin, Ziyou Wang, Lei Jin, Xin Zhang, Yinzhou Tang, Haisheng Su, et al. Worldarena: A unified benchmark for evaluating perception and functional utility of embodied world models. arXiv preprint arXiv:2602.08971, 2026

    work page arXiv 2026

  52. [52]

    informativeness

    Ansh Kumar Sharma, Yixiang Sun, Ninghao Lu, Y unzhe Zhang, Jiarao Liu, and Sherry Y ang. World-gymnast: Training robots with reinforcement learning in a world model. arXiv preprint arXiv:2602.02454, 2026

    work page arXiv 2026

  53. [53]

    Latent diffusion model without variational autoencoder

    Minglei Shi, Haolin Wang, Wenzhao Zheng, Ziyang Y uan, Xiaoshi Wu, Xintao Wang, Pengfei Wan, Jie Zhou, and Jiwen Lu. Latent diffusion model without variational autoencoder. In The F ourteenth Inter- national Conference on Learning Representations , 2026. URL https://openreview.net/forum? id=kdpeJNbFyf

    2026

  54. [54]

    Improving the diffusability of autoencoders

    Ivan Skorokhodov, Sharath Girish, Benran Hu, Willi Menapace, Y anyu Li, Rameen Abdal, Sergey Tulyakov, and Aliaksandr Siarohin. Improving the diffusability of autoencoders. In F orty-second In- ternational Conference on Machine Learning , 2025. URL https://openreview.net/forum?id= 2hEDcA7xy4

    2025

  55. [55]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent V anhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

    2016

  56. [56]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision , pages 402–419. Springer, 2020

    2020

  57. [57]

    Predictive inverse dynamics models are scalable learners for robotic manipulation

    Y ang Tian, Sizhe Y ang, Jia Zeng, Ping Wang, Dahua Lin, Hao Dong, and Jiangmiao Pang. Predictive inverse dynamics models are scalable learners for robotic manipulation. In The Thirteenth Interna- tional Conference on Learning Representations , 2025. URL https://openreview.net/forum?id= meRCKuUpmc

    2025

  58. [58]

    Mujoco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Y uval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems , pages 5026–5033. IEEE, 2012

    2012

  59. [59]

    Beyond language modeling: An exploration of multimodal pretraining.arXiv preprint arXiv:2603.03276,

    Shengbang Tong, David Fan, John Nguyen, Ellis Brown, Gaoyue Zhou, Shengyi Qian, Boyang Zheng, Théophane V allaeys, Junlin Han, Rob Fergus, et al. Beyond language modeling: An exploration of multimodal pretraining. arXiv preprint arXiv:2603.03276, 2026

    work page arXiv 2026

  60. [60]

    arXiv preprint arXiv:2601.16208 (2026),https://arxiv.org/abs/2601.16208

    Shengbang Tong, Boyang Zheng, Ziteng Wang, Bingda Tang, Nanye Ma, Ellis Brown, Jihan Y ang, Rob Fergus, Y ann LeCun, and Saining Xie. Scaling text-to-image diffusion transformers with representation autoencoders. arXiv preprint arXiv:2601.16208, 2026

    work page arXiv 2026

  61. [61]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdul- mohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Y e Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025

    work page internal anchor Pith review arXiv 2025

  62. [62]

    Scalable policy evaluation with video world models.arXiv preprint arXiv:2511.11520, 2025

    Wei-Cheng Tseng, Jinwei Gu, Qinsheng Zhang, Hanzi Mao, Ming-Y u Liu, Florian Shkurti, and Lin Y en-Chen. Scalable policy evaluation with video world models. arXiv preprint arXiv:2511.11520 , 2025

    work page arXiv 2025

  63. [63]

    Towards Accurate Generative Models of Video: A New Metric & Challenges

    Thomas Unterthiner, Sjoerd V an Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018

    work page internal anchor Pith review arXiv 2018

  64. [64]

    Score-based generative modeling in latent space.Advances in neural information processing systems , 34:11287–11302, 2021

    Arash V ahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space.Advances in neural information processing systems , 34:11287–11302, 2021. 14 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models

    2021

  65. [65]

    Attention is all you need

    Ashish V aswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  66. [66]

    Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen- Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, Abraham Lee, Kuan Fang, Chelsea Finn, and Sergey Levine

    Homer Rich Walke, Kevin Black, Tony Z. Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen- Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, Abraham Lee, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata v2: A dataset for robot learning at scale. In Jie Tan, Marc Tou- ssaint, and Kourosh Darvish, editors, Proceedings of The 7th Conference on Ro...

    2023

  67. [67]

    DDT: Decoupled diffusion Transformer.arXiv:2504.05741, 2025

    Shuai Wang, Zhi Tian, Weilin Huang, and Limin Wang. Ddt: Decoupled diffusion transformer. arXiv preprint arXiv:2504.05741, 2025

    work page arXiv 2025

  68. [68]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Y e, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal mod- els in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025

    work page internal anchor Pith review arXiv 2025

  69. [69]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing , 13(4):600–612, 2004

    2004

  70. [70]

    Learning interactive real-world simulators

    Sherry Y ang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In NeurIPS Workshop on Generalization in Planning, 2023

    2023

  71. [71]

    Reconstruction vs

    Jingfeng Y ao, Bin Y ang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  72. [72]

    An image is worth 32 tokens for reconstruction and generation

    Qihang Y u, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems, 37:128940–128966, 2024

    2024

  73. [73]

    arXiv preprint arXiv:2505.09694 (2025)

    Hu Y ue, Siyuan Huang, Y ue Liao, Shengcong Chen, Pengfei Zhou, Liliang Chen, Maoqing Y ao, and Guanghui Ren. Ewmbench: Evaluating scene, motion, and semantic quality in embodied world models. arXiv preprint arXiv:2505.09694, 2025

    work page arXiv 2025

  74. [74]

    Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan L

    Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Y ana Wei, Jiahao Wang, Vishal M Patel, Paul Pu Liang, et al. World-in-world: World models in a closed-loop world.arXiv preprint arXiv:2510.18135, 2025

    work page arXiv 2025

  75. [75]

    Reinforcing action policies by prophesying

    Jiahui Zhang, Ze Huang, Chun Gu, Zipei Ma, and Li Zhang. Reinforcing action policies by prophesying. arXiv preprint arXiv:2511.20633, 2025

    work page arXiv 2025

  76. [76]

    Rae-nwm: Navigation world model in dense visual representation space

    Mingkun Zhang, Wangtian Shen, Fan Zhang, Haijian Qin, Zihao Pei, and Ziyang Meng. Rae-nwm: Navigation world model in dense visual representation space. arXiv preprint arXiv:2603.09241, 2026

    work page arXiv 2026

  77. [77]

    The unreasonable ef- fectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable ef- fectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 586–595, 2018

    2018

  78. [78]

    Both semantics and reconstruction matter: Making rep- resentation encoders ready for text-to-image generation and editing

    Shilong Zhang, He Zhang, Zhifei Zhang, Chongjian Ge, Shuchen Xue, Shaoteng Liu, Mengwei Ren, Soo Y e Kim, Y uqian Zhou, Qing Liu, et al. Both semantics and reconstruction matter: Making rep- resentation encoders ready for text-to-image generation and editing. arXiv preprint arXiv:2512.17909 , 2025

    work page arXiv 2025

  79. [79]

    Diffusion Transformers with Representation Autoencoders

    Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with representa- tion autoencoders. arXiv preprint arXiv:2510.11690, 2025. 15 Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models

    work page internal anchor Pith review arXiv 2025

  80. [80]

    DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning

    Gaoyue Zhou, Hengkai Pan, Y ann LeCun, and Lerrel Pinto. DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning. In F orty-Second International Conference on Machine Learning, June 2025

    2025

Showing first 80 references.