Pith. sign in

REVIEW 3 major objections 4 minor 65 references

The paper claims that video world models can look good and follow instructions yet still lack inherent reactivity — the ability to infer unstated, scene-conditioned consequences — and supports this with a 1,474-case benchmark in which no mo

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 03:17 UTC pith:OXPV6X6U

load-bearing objection A serious, useful benchmark whose central split survives scrutiny; treat the World Reactivity metric as behavioral, not causal, and fix the NeoVerse disclosure before publication. the 3 major comments →

arxiv 2608.02603 v1 pith:OXPV6X6U submitted 2026-08-03 cs.CV

WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity

Yuxue Yang , Shuyao Shang , Jiahe Wang , Zitong Zhou , Liang Tan , Junhan Zeng , Ruizhi Li , Junyan Li
show 8 more authors
Yu Liu Xiao Yang Yong Li Jun Zhu Hongsheng Li Tieniu Tan Lue Fan Zhaoxiang Zhang
This is my paper
classification cs.CV
keywords world modelsvideo generationinherent reactivitybenchmarkscene-conditioned reactionscontrol adherenceinterface adaptationevaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that evaluating video generation models as world models requires a fourth axis beyond visual quality, instruction following, and spatial memory: inherent reactivity, or the ability to infer from an initial scene what should happen next when it is not stated. To test this, it builds WorldExam, a benchmark of 1,474 cases across eight tasks, and evaluates 20 models through three control interfaces. The results reveal a capability split: camera-driven models control the viewpoint but cannot interact, action-driven models move subjects but leave the world unresponsive, and language-driven models handle interaction but follow complex controls poorly. The central claim is that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity, so world models need a separate reactivity diagnostic.

Core claim

WorldExam establishes that inherent reactivity is a distinct, measurable capability that no current video world model combines with broad task coverage. By adapting shared atomic control units (movement, pan/tilt, stop) to each model's native interface, the benchmark can compare camera-, action-, and language-driven models on the same 1,474 cases. On the World Reactivity level, the expected scene-conditioned consequences are deliberately left unstated in the input, so a high score requires the model to infer them from the scene state. Across 20 models, the paper finds that the best camera-control models do not support dynamic interaction; the best action-control models often generate static,

What carries the argument

The central mechanism is the atomic control unit: a small set of discrete camera and subject actions (forward/back/left/right, tilt/pan, stop) that WorldExam composes into ordered sequences and then adapts into each paradigm's native format—SE(3) camera trajectories, discrete action inputs, or natural-language prompts. This interface adaptation allows the same case to be posed to all three model families. On the evaluation side, the paper pairs geometry-based scoring (reconstructing camera, subject, and terrain trajectories) with checklist-based scoring: case-specific checklists written by an LLM, covering initiating condition, reaction, and invalid outcomes, are scored by a vision-language

Load-bearing premise

The benchmark's reactivity scores are only as valid as the LLM-written checklists and the vision-language judge's binary calls on 10 sampled frames; if these checklists reward surface heuristics ("contact happened, object moved") rather than true scene-conditioned causation, a model could score high without inferring consequences.

What would settle it

Take a high-scoring model and run a counterfactual: generate the same scene twice, once with a movable obstacle in the subject's path and once with it removed, keeping the control identical. If the model produces essentially the same reaction (or none) in both cases, its World Reactivity score reflects a surface heuristic rather than scene-conditioned inference; a model with genuine reactivity should alter its behavior when the obstacle is present.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity, then visual-quality scores (e.g., subject consistency, aesthetic quality) should not be used as proxies for world-model capability; benchmarks should report reactivity separately.
  • Since camera-driven models cannot handle dynamic interaction, any unified world-model claim from a camera-only model is incomplete; interaction support must be part of the specification.
  • Action-driven models that precisely control a subject while leaving terrain, objects, and agents unresponsive indicate that control adherence is disjoint from scene-conditioned inference; improving one will not automatically improve the other.
  • Language-driven models' weaker control adherence suggests that natural-language interfaces need to be paired with explicit control structure (e.g., temporal segmentation) to be comparable to direct action interfaces.
  • The two-track design implies that interface-compatible evaluation, not a single global ranking, is the right way to compare heterogeneous world models without penalizing unsupported capabilities.

Where Pith is reading between the lines

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

  • The diagnostic levels may generalize to other generative domains: any system that claims to simulate a world (e.g., robotics simulators, game engines) could be evaluated on the same inherent-reactivity axis by withholding expected outcomes and checking inferred consequences.
  • A testable extension: rather than scoring a single generated rollout, intervention tests—perturbing an obstacle, an agent, or a physical property and comparing reactions across rollouts—would provide stronger evidence of scene-conditioned inference than any single-video checklist.
  • If reactivity is truly absent in current models, the bottleneck may not be visual synthesis but the lack of a learned physics or interaction prior; future models might need explicit scene graphs or dynamics modules rather than scaling generation alone.
  • The paper's human-alignment study validates the judge but not the checklist construct; a follow-up could ask humans to rate whether each checklist item genuinely tests scene-conditioned causality versus surface plausibility, which would tighten the benchmark's central construct.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. WorldExam is a hierarchical benchmark for controllable video generation models, with 1,474 cases spanning four diagnostic levels (Visual Quality, Control Adherence, Spatial Consistency, World Reactivity) and eight tasks. The benchmark adapts shared control intents to camera-, action-, and language-driven interfaces using atomic control units. Twenty representative models are evaluated on static-scene and dynamic-interaction tracks. The central empirical finding is a clear capability split: camera-driven models are strongest at camera control but lack interaction support; action-driven models control subjects relatively well but often fail to produce scene-conditioned reactions; language-driven models do better on interaction and goal completion but follow composed controls less faithfully. The paper concludes that no model combines broad task coverage with consistently strong performance, so high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity. The paper also reports a human-alignment study for the checklist-based tasks and a backend-stability analysis with an alternative depth reconstruction model.

Significance. If the benchmark and its central claim hold, WorldExam makes a valuable contribution by separating explicit instruction fulfillment from scene-conditioned reactivity, a distinction that most existing benchmarks do not directly address. The benchmark is relatively large, covers diverse scenes, interfaces, and tasks, and includes a sensible separation of tracks by interface capabilities. The human-alignment study (Section 5.6) and the backend-stability check (Section 5.7) are praiseworthy empirical safeguards: they show that the VLM judge agrees with human raters under the same protocol and that the main rankings are not highly sensitive to the reconstruction backend. The central capability split is supported by the reported data, e.g., Kling 2.5's high General average (81.04) versus low Goal Completion (48.25), and the sharp drop for action-driven models from Subject Control to Terrain Interaction. However, the construct validity of the World Reactivity metric is the main open question, because the checklists are LLM-authored and the VLM judge may share statistical priors with language-driven models. This tempers the strength of the 'inherent reactivity' interpretation, although the

major comments (3)
  1. [§4.3, §3.3, §5.6] The World Reactivity score is the fraction of checklist items judged satisfied by a VLM on 10 uniformly sampled frames (Eq. 4). The human-alignment study in Section 5.6 (Spearman ρ=0.86) establishes agreement between human raters and the VLM applying the same checklists to the same frames. This is inter-rater reliability for the protocol, not construct validity: it does not show that the checklists capture scene-conditioned inference rather than surface-plausibility heuristics such as 'contact happens; object moves.' Because the checklists are authored by an LLM (Section 3.3) and many tested models are language-driven, part of the language-driven advantage may reflect a shared statistical prior with the evaluator. The paper itself acknowledges in Section 6 that the metrics assess observable end-to-end behavior and do not establish internal causal representation. The abstract's claim of '
  2. [§4.2, Eq. (5)] The image-space displacement calibration factor k_{m,c}=W/(2d_{m,c}) is computed from the model's own default-input displacement d_{m,c}. If a camera-driven model follows camera control poorly, d_{m,c} is small, so k_{m,c} is large, which increases the input translation magnitude used in the final evaluation. The ablation in Table 3 shows that larger translation multipliers degrade Camera Control, Scene Revisit, and general metrics for NeoVerse. This creates an endogeneity concern: models with weak camera control may receive systematically harder inputs, potentially exaggerating differences among camera-driven models. Please report the distribution of k_{m,c} across models and cases, and test whether the camera-driven ranking is stable when using a fixed translation multiplier or an externally specified target displacement.
  3. [§5.1, Table 4] The dynamic-interaction track for action-driven models includes only two models (WorldPlay and LingBot-World) because the other five action-driven models lack reliable third-person subject control. The claim that action-driven models 'control subjects more precisely but often leave the world unresponsive' is therefore based on a very small sample. While the eligibility restriction is defensible and clearly stated, the conclusion should be phrased more cautiously, and the paper should acknowledge that the action-driven row of the capability split rests on only two systems.
minor comments (4)
  1. [Table 2 / Refs] The paper evaluates the authors' own model NeoVerse [56] on the static-scene track, where it achieves the highest Overall score (85.39). This is not disclosed in the main text. The authors should explicitly state this relationship in the experimental setup or in a footnote, per standard transparency practice for benchmark papers.
  2. [§4.3 / Appendix C] Several checklist items ask whether the relevant objects/agents 'remain visible enough to judge' or whether unrelated dominant events occur. These items may reward conservative, low-event videos and could interact with the 10-frame subsampling. A short discussion of how such items affect the final scores would improve interpretability.
  3. [§4.1] For language-driven models, Camera Control segments are inferred by change-point detection and matched to atomic controls in temporal order. This is a reasonable approach, but it can silently misalign segments if the model performs the controls in a different order or merges them. A brief sensitivity analysis of the change-point detection parameters would strengthen the camera-control comparison.
  4. [General] The paper uses many threshold values (e.g., 10% of maximum displacement for Scene Revisit, 0.5 translation error, 15° rotation error). These thresholds appear reasonable, but their choice is not justified. Since they directly affect pass/fail scores, a sentence citing a principled basis or reporting sensitivity to threshold variation would be useful.

Circularity Check

0 steps flagged

No circularity: the central claims are empirical benchmark observations, not derivations; self-citation and construct-validity limitations do not reduce outputs to inputs.

full rationale

WorldExam is an evaluation benchmark, not a derivation. The central claim—that no model combines broad task coverage with strong performance—is an empirical aggregate of scores from 20 models, not a quantity entailed by the metric definitions. The World Reactivity score is an operationalization (Eq. 4: fraction of checklist items satisfied), and the checklist items are withheld from the model input; there is no fitted parameter renamed as a prediction, and no equation in the paper reduces a predicted outcome to an input by construction. The self-evaluation of NeoVerse [56] is a transparency concern, but it is not load-bearing: NeoVerse participates only in the static-scene track, while the capability-split and reactivity conclusions come from the dynamic-interaction track where camera-driven models are excluded. The paper's own §6 limitation—'the metrics assess observable end-to-end behavior rather than determining where reasoning occurs or establishing that the video generator itself has learned an internal causal representation'—is an honest construct-validity caveat, not evidence of circularity. The human-alignment study (Tab. 5) validates the VLM judge against human raters applying the same checklists to the same frames; that is inter-rater reliability for the protocol, not a self-referential proof of the construct. No uniqueness theorem or ansatz is imported from the authors' prior work, and no renaming of a known result is presented as a derivation. The benchmark is self-contained in its evaluation, and the stated limitations are acknowledged rather than hidden.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The evaluation rests on unproven background premises: monocular 3D reconstruction accuracy, VLM-judge reliability, cross-interface comparability of the control adaptation, and LLM-authored checklist validity. The paper partially tests the first two (DA3 backend stability; human alignment rho=0.86) but does not directly test the latter two. Calibration factors and metric thresholds are hand-fitted design choices, not fitted to the conclusion. No new physical entities are postulated.

free parameters (5)
  • Camera translation calibration factor k_{m,c} = W/(2*d_{m,c}) per model-case
    Fitted in the Section 4.2 calibration pass from SAM2-tracked anchor displacement; normalizes image-space displacement across camera-driven models. It changes input difficulty per model but is transparent and pre-registered in the protocol.
  • Camera Control metric thresholds = e_t,max=0.5; e_r,max=15 deg; segment min displacement 5%; rotation min 5 deg
    Hand-chosen caps and normalizers in Section 4.1 for translation/rotation errors; arbitrary but stated, and they affect the Camera Control score scale.
  • Scene Revisit thresholds and fractions = success: translation <=10% of max displacement, rotation <5 deg; rho_1=0.4, rho_2=0.6; PSNR/25 cap
    Design choices in Section 4.1 defining revisit success, execution-time allocation, and appearance normalization.
  • Subject Control / Terrain Interaction thresholds = e_t,max=0.5; min displacement 0.5% of scene scale
    Hand-set thresholds in Section 4.3; Subject Control score S_sub = 100*max(0, 1-e_t/0.5).
  • Execution-time fractions for static-scene cases = per-case rho_i; Scene Revisit 0.4/0.6
    Case-construction choices in Sections 3.2 and 4.1 that affect task difficulty and segmentation.
axioms (5)
  • domain assumption Monocular reconstruction (VGGT-Omega or DA3) recovers camera/subject trajectories and scene geometry accurately enough for the metrics' thresholds, and the gravity direction from the horizontal-region mask is correct.
    Used in Sections 4.1 and 4.3 for Camera Control, Scene Revisit, Subject Control, and Terrain Interaction; tested only against one alternative reconstruction backend (DA3, Section 5.7).
  • domain assumption The 10 uniformly sampled frames plus the case-specific checklist carry enough information for a VLM (GPT-5.5) to judge checklist items, and humans reviewing the same frames agree (rho=0.86, Section 5.6).
    Checklist evaluation in Section 4.3; validation is judge-vs-human on the same frames, not construct validity of the checklists against an external standard of 'reactivity'.
  • ad hoc to paper Interface adaptation via atomic control units is a faithful equivalence across camera-, action-, and language-driven inputs.
    Defined in Sections 3 and 4.1; no validation that the three adapted instructions produce equal task difficulty, so cross-paradigm comparisons rest on this premise.
  • ad hoc to paper LLM-generated initial scenarios, after human filtering, are valid pre-event states in which the intended reaction is physically feasible and unambiguous.
    Section 3.3 curation pipeline; human filtering discards infeasible cases, but feasibility is judged by the authors' criteria.
  • domain assumption For API-based commercial models, proprietary prompt enhancement is allowed as part of the model pipeline.
    Section 5.1 states this; it confounds 'model capability' with the vendor's prompt processing when interpreting paradigm-level results.
invented entities (1)
  • WorldExam benchmark (1,474 cases, eight tasks, four diagnostic levels) independent evidence
    purpose: Standardized diagnostic instrument for evaluating video world models on apparent appearance and inherent reactivity.
    A benchmark is a measurement instrument rather than a physical or theoretical entity; the scores it yields are falsifiable because model outputs are checkable, so it carries an external handle. Listed for completeness of the ledger.

pith-pipeline@v1.3.0-daily-deepseek · 25946 in / 20844 out tokens · 754250 ms · 2026-08-04T03:17:12.257295+00:00 · methodology

0 comments
read the original abstract

Controllable video generation models are increasingly being developed as world models. Accordingly, evaluating them in this role extends beyond the apparent appearance of generated videos to the inherent reactivity of the worlds they depict: the ability to infer from the scene state how the world should react and to generate plausible consequences not explicitly described in the input. Yet existing benchmarks mainly assess visual quality or explicit instruction fulfillment by checking whether requested actions and interaction outcomes are realized, leaving inherent reactivity underexamined. We introduce WorldExam, a hierarchical diagnostic benchmark spanning four levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. It comprises 1,474 cases across eight dedicated tasks and supports unified evaluation of camera-, action-, and language-driven model paradigms. The World Reactivity level evaluates scene-conditioned reactions and goal-directed behaviors beyond what is explicitly specified in the input. Evaluation of 20 representative models reveals a clear capability split. Camera-driven models excel at camera control, but their interfaces do not support dynamic interaction; action-driven models control subjects more precisely but often leave the world unresponsive; and language-driven models perform better on interaction but follow complex controls less faithfully. No model combines broad task coverage with consistently strong performance, showing that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity.

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

65 extracted references · 30 linked inside Pith

  1. [1]

    HappyHorse 1.0 I2V, 2026

    Alibaba Group. HappyHorse 1.0 I2V, 2026. URL https://www.alibabacloud.com/help/en/model-studio/ happyhorse-image-to-video-api-reference

    2026

  2. [2]

    Act-bench: Towards action controllable world models for autonomous driving.arXiv preprint arXiv:2412.05337, 2024

    Hidehisa Arai, Keishi Ishihara, Tsubasa Takahashi, and Yu Yamaguchi. Act-bench: Towards action controllable world models for autonomous driving.arXiv preprint arXiv:2412.05337, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Recammaster: Camera-controlled generative rendering from a single video

    Jianhong Bai, Menghan Xia, Xiao Fu, Xintao Wang, Lianrui Mu, Jinwen Cao, Zuozhu Liu, Haoji Hu, Xiang Bai, Pengfei Wan, et al. Recammaster: Camera-controlled generative rendering from a single video. InICCV, pages 14834–14844. IEEE, 2025

    2025

  4. [4]

    Videophy: Evaluating physical commonsense for video generation

    Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. In ICLR, volume 2025, pages 102075–102121, 2025

    2025

  5. [5]

    Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models

    Fan Bao, Chendong Xiang, Gang Yue, Guande He, Hongzhou Zhu, Kaiwen Zheng, Min Zhao, Shilong Liu, Yaole Wang, and Jun Zhu. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv preprint arXiv:2405.04233, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Video generation models as world simulators.OpenAI Blog, 1(8):1, 2024

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Leo Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators.OpenAI Blog, 1(8):1, 2024

    2024

  7. [7]

    Genie: Generative interactive environments

    Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. InForty-first international conference on machine learning, 2024

    2024

  8. [8]

    T2vworldbench: A benchmark for evaluating world knowledge in text-to-video generation

    Yubin Chen, Xuyang Guo, Zhenmei Shi, Zhao Song, and Jiahao Zhang. T2vworldbench: A benchmark for evaluating world knowledge in text-to-video generation. InWACV, pages 6474–6485. IEEE, 2026

    2026

  9. [9]

    Fantasyworld: Geometry-consistent world modeling via unified video and 3d prediction

    Yixiang Dai, Fan Jiang, Chiyu Wang, Mu Xu, and Yonggang Qi. Fantasyworld: Geometry-consistent world modeling via unified video and 3d prediction. InICLR, volume 2026, pages 103603–103622, 2026

    2026

  10. [10]

    Worldscore: A unified evaluation benchmark for world generation

    Haoyi Duan, Hong-Xing Yu, Sirui Chen, Li Fei-Fei, and Jiajun Wu. Worldscore: A unified evaluation benchmark for world generation. InICCV, pages 27713–27724, 2025

    2025

  11. [11]

    iworld-bench: A benchmark for interactive world models with a unified action generation framework.arXiv preprint arXiv:2605.03941, 2026

    Jianjie Fang, Yingshan Lei, Qin Wan, Ziyou Wang, Yuchao Huang, Yongyan Xu, Baining Zhao, Weichen Zhang, Chen Gao, Xinlei Chen, et al. iworld-bench: A benchmark for interactive world models with a unified action generation framework.arXiv preprint arXiv:2605.03941, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Veo 3 technical report

    Google DeepMind. Veo 3 technical report. Technical report, Google DeepMind, 2025. URLhttps://storage. googleapis.com/deepmind-media/veo/Veo-3-Tech-Report.pdf

    2025

  13. [13]

    Voyager: Long-range and world-consistent video diffusion for explorable 3d scene generation.ACM TOG, 44(6):1–15, 2025

    Tianyu Huang, Wangguandong Zheng, Tengfei Wang, Yuhao Liu, Zhenwei Wang, Junta Wu, Jie Jiang, Hui Li, Rynson Lau, Wangmeng Zuo, et al. Voyager: Long-range and world-consistent video diffusion for explorable 3d scene generation.ACM TOG, 44(6):1–15, 2025

    2025

  14. [14]

    Vbench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. InCVPR, pages 21807–21818. IEEE, 2024

    2024

  15. [15]

    Vbench++: Comprehensive and versatile benchmark suite for video generative models

    Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, et al. Vbench++: Comprehensive and versatile benchmark suite for video generative models. TPAMI, 2025

    2025

  16. [16]

    Robowm-bench: A benchmark for evaluating world models in robotic manipulation.arXiv preprint arXiv:2604.19092, 2026

    Feng Jiang, Yang Chen, Kyle Xu, Yuchen Liu, Haifeng Wang, Zhenhao Shen, Jasper Lu, Shengze Huang, Yuanfei Wang, Chen Xie, et al. Robowm-bench: A benchmark for evaluating world models in robotic manipulation.arXiv preprint arXiv:2604.19092, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Kling AI.https://klingai.com, 2025

    Kuaishou Technology. Kling AI.https://klingai.com, 2025. URLhttps://klingai.com

    2025

  18. [18]

    Worldmodelbench: Judging video generation models as world models

    Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu Yin, Joseph Gonzalez, et al. Worldmodelbench: Judging video generation models as world models. NeurIPS, 38, 2026. 24

    2026

  19. [19]

    Robotrustbench: Benchmarking the trustworthiness of video world models for robotic manipulation.arXiv preprint arXiv:2606.01600, 2026

    Huiqiong Li, Jiayu Wang, Zhiting Mei, Anirudha Majumdar, Jingjing Chen, and Bin Zhu. Robotrustbench: Benchmarking the trustworthiness of video world models for robotic manipulation.arXiv preprint arXiv:2606.01600, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Hunyuan-gamecraft: High-dynamic interactive game video generation with hybrid history condition

    Jiaqi Li, Junshu Tang, Zhiyong Xu, Longhuang Wu, Yuan Zhou, Shuai Shao, Tianbao Yu, Zhiguo Cao, and Qinglin Lu. Hunyuan-gamecraft: High-dynamic interactive game video generation with hybrid history condition. arXiv preprint arXiv:2506.17201, 2025

    Pith/arXiv arXiv 2025

  21. [21]

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

    Yaxuan 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

    Pith/arXiv arXiv 2025

  22. [22]

    Worldlens: Full-spectrum evaluations of driving world models in real world

    Ao Liang, Lingdong Kong, Tianyi Yan, Hongsi Liu, Yu Yang, Ziqi Huang, Wei Yin, Jialong Zuo, Yixuan Hu, Dekai Zhu, et al. Worldlens: Full-spectrum evaluations of driving world models in real world. InCVPR, pages 36385–36399, 2026

    2026

  23. [23]

    Depth anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647, 2025

    Haotong Lin, Sili Chen, Junhao Liew, Donny Y Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Phyground: Benchmarking physical reasoning in generative world models.arXiv preprint arXiv:2605.10806, 2026

    Juyi Lin, Arash Akbari, Yumei He, Lin Zhao, Haichao Zhang, Arman Akbari, Xingchen Xu, Zoe Y Lu, Enfu Nan, Hokin Deng, et al. Phyground: Benchmarking physical reasoning in generative world models.arXiv preprint arXiv:2605.10806, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InCVPR, pages 22160–22169. IEEE, 2024

    2024

  26. [26]

    Rise-video: Can video generators decode implicit world rules?arXiv preprint arXiv:2602.05986, 2026

    Mingxin Liu, Shuran Ma, Shibei Meng, Xiangyu Zhao, Zicheng Zhang, Shaofeng Zhang, Zhihang Zhong, Peixian Chen, Haoyu Cao, Xing Sun, et al. Rise-video: Can video generators decode implicit world rules?arXiv preprint arXiv:2602.05986, 2026

    arXiv 2026

  27. [27]

    Evalcrafter: Benchmarking and evaluating large video generation models

    Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In CVPR, pages 22139–22149. IEEE, 2024

    2024

  28. [28]

    Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation.NeurIPS, 36:62352–62387, 2023

    Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation.NeurIPS, 36:62352–62387, 2023

    2023

  29. [29]

    Kinebench: Benchmarking embodied world models via idm-free kinematic grounding.arXiv preprint arXiv:2607.19876, 2026

    Zeyu Liu, Zhangzhe Zhu, Yang Zhang, Chenyou Fan, Chenjia Bai, and Xuelong Li. Kinebench: Benchmarking embodied world models via idm-free kinematic grounding.arXiv preprint arXiv:2607.19876, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    Xiaofeng Mao, Zhen Li, Chuanhao Li, Xiaojie Xu, Kaining Ying, and Kaipeng Zhang. Yume1. 5: A text-controlled interactive world generation model. InCVPR, pages 7752–7761, 2026

    2026

  31. [31]

    Towards world simulator: Crafting physical commonsense-based benchmark for video generation

    Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Hailuo AI Video.https://hailuoai.video/, 2024

    MiniMax. Hailuo AI Video.https://hailuoai.video/, 2024. URLhttps://hailuoai.video/

    2024

  33. [33]

    Worldsimbench: Towards video generation models as world simulators.arXiv preprint arXiv:2410.18072, 2024

    Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, Lu Sheng, Jing Shao, et al. Worldsimbench: Towards video generation models as world simulators.arXiv preprint arXiv:2410.18072, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. InICLR, volume 2025, pages 28085–28128, 2025

    2025

  35. [35]

    Seedance 2.0: Advancing video generation for world complexity.arXiv preprint arXiv:2604.14148, 2026

    Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, Yufeng Cheng, et al. Seedance 2.0: Advancing video generation for world complexity.arXiv preprint arXiv:2604.14148, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    Worldarena: A unified benchmark for evaluating perception and functional utility of embodied world models.arXiv preprint arXiv:2602.08971, 2026

    Yu 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

    arXiv 2026

  37. [37]

    T2v-compbench: A comprehensive benchmark for compositional text-to-video generation

    Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. InCVPR, pages 8406–8416. IEEE, 2025. 25

    2025

  38. [38]

    Worldplay: Towards long-term geometric consistency for real-time interactive world modeling.arXiv preprint arXiv:2512.14614, 2025

    Wenqiang Sun, Haiyu Zhang, Haoyuan Wang, Junta Wu, Zehan Wang, Zhenwei Wang, Yunhong Wang, Jun Zhang, Tengfei Wang, and Chunchao Guo. Worldplay: Towards long-term geometric consistency for real-time interactive world modeling.arXiv preprint arXiv:2512.14614, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Inspatio-world: A real-time 4d world simulator via spatiotemporal autoregressive modeling.arXiv preprint arXiv:2604.07209, 2026

    InSpatio Team, Donghui Shen, Guofeng Zhang, Haomin Liu, Haoyu Ji, Hujun Bao, Hongjia Zhai, Jialin Liu, Jing Guo, Nan Wang, et al. Inspatio-world: A real-time 4d world simulator via spatiotemporal autoregressive modeling.arXiv preprint arXiv:2604.07209, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Advancing open-source world models.arXiv preprint arXiv:2601.20540, 2026

    Robbyant Team, Zelin Gao, Qiuyu Wang, Yanhong Zeng, Jiapeng Zhu, Ka Leong Cheng, Yixuan Li, Hanlin Wang, Yinghao Xu, Shuailei Ma, et al. Advancing open-source world models.arXiv preprint arXiv:2601.20540, 2026

    Pith/arXiv arXiv 2026

  41. [41]

    Ntire 2017 challenge on single image super-resolution: Methods and results

    Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. InCVPRW, pages 1110–1121. IEEE, 2017

    2017

  42. [42]

    Selective review of offline change point detection methods

    Charles Truong, Laurent Oudre, and Nicolas Vayatis. Selective review of offline change point detection methods. Signal processing, 167:107299, 2020

    2020

  43. [43]

    Worldbench: Disambiguating physics for diagnostic evaluation of world models.arXiv preprint arXiv:2601.21282, 2026

    Rishi Upadhyay, Howard Zhang, Jim Solomon, Ayush Agrawal, Pranay Boreddy, Shruti Satya Narayana, Yunhao Ba, Alex Wong, Celso M de Melo, and Achuta Kadambi. Worldbench: Disambiguating physics for diagnostic evaluation of world models.arXiv preprint arXiv:2601.21282, 2026

    arXiv 2026

  44. [44]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Vggt-ω.arXiv preprint arXiv:2605.15195, 2026

    Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. Vggt-ω.arXiv preprint arXiv:2605.15195, 2026

    Pith/arXiv arXiv 2026

  46. [46]

    Matrix-game 3.0: Real-time and streaming interactive world model with long-horizon memory

    Zile Wang, Zexiang Liu, Jiaxing Li, Kaichen Huang, Baixin Xu, Fei Kang, Mengyin An, Peiyu Wang, Biao Jiang, Yichen Wei, et al. Matrix-game 3.0: Real-time and streaming interactive world model with long-horizon memory. arXiv preprint arXiv:2604.08995, 2026

    Pith/arXiv arXiv 2026

  47. [47]

    Quantitative video world model evaluation for geometric-consistency.arXiv preprint arXiv:2605.15185, 2026

    Jiaxin Wu, Yihao Pi, Yinling Zhang, Yuheng Li, and Xueyan Zou. Quantitative video world model evaluation for geometric-consistency.arXiv preprint arXiv:2605.15185, 2026

    Pith/arXiv arXiv 2026

  48. [48]

    Worldreasonbench: Human-aligned stress testing of video generators as future world-state predictors.arXiv preprint arXiv:2605.10434, 2026

    Keming Wu, Yijing Cui, Wenhan Xue, Qijie Wang, Xuan Luo, Zhiyuan Feng, Zuhao Yang, Sudong Wang, Sicong Jiang, Haowei Zhu, et al. Worldreasonbench: Human-aligned stress testing of video generators as future world-state predictors.arXiv preprint arXiv:2605.10434, 2026

    Pith/arXiv arXiv 2026

  49. [49]

    Omni-worldbench: Towards a comprehensive interaction-centric evaluation for world models.arXiv preprint arXiv:2603.22212, 2026

    Meiqi Wu, Zhixin Cai, Fufangchen Zhao, Xiaokun Feng, Rujing Dang, Bingze Song, Ruitian Tian, Jiashu Zhu, Jiachen Lei, Hao Dou, et al. Omni-worldbench: Towards a comprehensive interaction-centric evaluation for world models.arXiv preprint arXiv:2603.22212, 2026

    arXiv 2026

  50. [50]

    Infinite-world: Scaling interactive world models to 1000-frame horizons via pose-free hierarchical memory.arXiv preprint arXiv:2602.02393, 2026

    Ruiqi Wu, Xuanhua He, Meng Cheng, Tianyu Yang, Yong Zhang, Zhuoliang Kang, Xunliang Cai, Xiaoming Wei, Chunle Guo, Chongyi Li, et al. Infinite-world: Scaling interactive world models to 1000-frame horizons via pose-free hierarchical memory.arXiv preprint arXiv:2602.02393, 2026

    arXiv 2026

  51. [51]

    Worldroambench: An open-world benchmark for long-horizon stability of interactive world models.arXiv preprint arXiv:2606.31672, 2026

    Ting-Bing Xu, Jiacheng Sui, Zhe Gao, Kewei Shi, Wenjin Yang, Zhicheng Liu, Zhaoxu Sun, Mingchao Sun, Hongyu Pan, Fan Jiang, et al. Worldroambench: An open-world benchmark for long-horizon stability of interactive world models.arXiv preprint arXiv:2606.31672, 2026

    Pith/arXiv arXiv 2026

  52. [52]

    Worldmark: A unified benchmark suite for interactive video world models.arXiv preprint arXiv:2604.21686, 2026

    Xiaojie Xu, Zhengyuan Lin, Kang He, Yukang Feng, Xiaofeng Mao, Yuanyang Yin, Kaipeng Zhang, and Yongtao Ge. Worldmark: A unified benchmark suite for interactive video world models.arXiv preprint arXiv:2604.21686, 2026

    Pith/arXiv arXiv 2026

  53. [53]

    Acwm- phys: Investigating generalized physical interaction in action-conditioned video world models.arXiv preprint arXiv:2605.08567, 2026

    Haotian Xue, Yipu Chen, Liqian Ma, Zelin Zhao, Lama Moukheiber, Yuchen Zhu, and Yongxin Chen. Acwm- phys: Investigating generalized physical interaction in action-conditioned video world models.arXiv preprint arXiv:2605.08567, 2026

    Pith/arXiv arXiv 2026

  54. [54]

    Learning interactive real-world simulators.arXiv preprint arXiv:2310.06114, 2023

    Sherry Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Tompson, Leslie Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators.arXiv preprint arXiv:2310.06114, 2023

    Pith/arXiv arXiv 2023

  55. [55]

    Mirabench: Evaluating action-conditioned reliability in robotic world models.arXiv preprint arXiv:2605.29360, 2026

    Tianzhuo Yang, Zihan Shen, Zirui Mi, Zhaoyi Zhang, Jiayi Zhou, Jiaming Ji, Juntao Dai, Jiawei Chen, Boyuan Chen, and Yaodong Yang. Mirabench: Evaluating action-conditioned reliability in robotic world models.arXiv preprint arXiv:2605.29360, 2026. 26

    Pith/arXiv arXiv 2026

  56. [56]

    Neoverse: Enhancing 4d world model with in-the-wild monocular videos.arXiv preprint arXiv:2601.00393, 2026

    Yuxue Yang, Lue Fan, Ziqi Shi, Junran Peng, Feng Wang, and Zhaoxiang Zhang. Neoverse: Enhancing 4d world model with in-the-wild monocular videos.arXiv preprint arXiv:2601.00393, 2026

    arXiv 2026

  57. [57]

    Mind: Benchmarking memory consistency and action control in world models.arXiv preprint arXiv:2602.08025, 2026

    Yixuan Ye, Xuanyu Lu, Yuxin Jiang, Yuchao Gu, Rui Zhao, Qiwei Liang, Jiachun Pan, Fengda Zhang, Weijia Wu, and Alex Jinpeng Wang. Mind: Benchmarking memory consistency and action control in world models.arXiv preprint arXiv:2602.08025, 2026

    arXiv 2026

  58. [58]

    Wbench: A comprehensive multi-turn benchmark for interactive video world model evaluation

    Kaining Ying, Hengrui Hu, Siyu Ren, Jiamu Li, Fengjiao Chen, Ziwen Wang, Xuezhi Cao, Xunliang Cai, and Henghui Ding. Wbench: A comprehensive multi-turn benchmark for interactive video world model evaluation. arXiv preprint arXiv:2605.25874, 2026

    Pith/arXiv arXiv 2026

  59. [59]

    Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models

    Mark Yu, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. InICCV, pages 100–111. IEEE, 2025

    2025

  60. [60]

    Ewmbench: Evaluating scene, motion, and semantic quality in embodied world models.arXiv preprint arXiv:2505.09694, 2025

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

    Pith/arXiv arXiv 2025

  61. [61]

    World-in-world: World models in a closed-loop world

    Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Jiahao Wang, Vishal Patel, Paul Liang, et al. World-in-world: World models in a closed-loop world. InICLR, volume 2026, pages 55660–55699, 2026

    2026

  62. [62]

    Worldolympiad: Can your world model survive a triathlon?arXiv preprint arXiv:2606.11129, 2026

    Yuke Zhao, Wangbo Zhao, Weijie Wang, Zeyu Zhang, Dakai An, Akide Liu, Yinghao Yu, Jiasheng Tang, Fan Wang, Wei Wang, et al. Worldolympiad: Can your world model survive a triathlon?arXiv preprint arXiv:2606.11129, 2026

    Pith/arXiv arXiv 2026

  63. [63]

    Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025

    Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Lulu Gu, Yuanhan Zhang, Jingwen He, Wei-Shi Zheng, et al. Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025

    Pith/arXiv arXiv 2025

  64. [64]

    Drivinggen: A comprehensive benchmark for generative video world models in autonomous driving

    Yang Zhou, Hao Shao, Letian Wang, Zhuofan Zong, Hongsheng Li, and Steven Waslander. Drivinggen: A comprehensive benchmark for generative video world models in autonomous driving. InICLR, volume 2026, pages 103502–103524, 2026

    2026

  65. [65]

    Astra: General interactive world model with autoregressive denoising

    Yixuan Zhu, Jiaqi Feng, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Jiwen Lu, and Jie Zhou. Astra: General interactive world model with autoregressive denoising. InICLR, volume 2026, pages 79167–79184, 2026. 27

    2026