Sources · Updated 2026-08-21
Sources
An article source map: see which publications support the site's explanations, what role each source plays and where the evidence remains limited. The full verification method is explained separately.
这是本站的文章来源地图：集中展示各篇解释依赖哪些资料、每类来源承担什么作用，以及证据仍有哪些限制。完整核验逻辑另见核验标准。
Short answer / 一句话结论
This page is frozen as of 2026-08-21. The source and claim records below are no longer rechecked on a schedule. Each was checked on the date it names, and those dates are real; none of them says anything about what a source contains today. Nothing has been deleted, and errors are still corrected when noticed or reported. See For agents for the full freeze notice.
本页已于 2026-08-21 冻结。以下来源与判断记录不再做周期性复查。 每条记录都在它标注的日期核对过，日期为真；但它不说明来源今天的内容。历史记录一条不删，发现或收到纠错时仍会修改。 完整冻结说明见 For agents。
Sources on this site are evidence inputs, not automatic truth. Official claims stay labeled as official claims unless independent validation is available.
本站把来源当作证据输入，而不是自动真相。官方声明会继续标成官方声明，除非有独立验证或更强证据支持。
Technical mappings: article, team and claim IDs / 技术映射：文章、团队与判断 ID
Article source map / 文章来源映射
Article pages are the human-facing explanations. Their evidence tables now expose claim IDs that match claims.json.
文章页负责给人读的解释。每篇文章的证据表现在会显示对应的 claim ID，并与 claims.json 保持一致。
| Article | Primary use | Main claim IDs | Main source IDs |
|---|---|---|---|
| What is a world model? | Core definition, prediction, action, internal simulation. |
claim-world-model-dynamics,
claim-video-generation-world-simulators,
claim-genie-3-interactive-worlds,
claim-vjepa-2-predictive-representation
|
world-models-2018,
openai-sora-world-simulators,
deepmind-genie-3,
meta-vjepa-2
|
| Is Sora a world model? | Boundary between video generation and interactive world modeling. |
claim-video-generation-world-simulators,
claim-sora-minute-video,
claim-sora-world-like-properties,
claim-sora-not-full-public-interactive-world-model
|
openai-sora-world-simulators,
world-models-2018,
deepmind-genie-3
|
| Genie, Cosmos and V-JEPA | Comparison of interactive worlds, physical-AI infrastructure and predictive representations. |
claim-genie-3-interactive-worlds,
claim-genie-3-public-performance,
claim-cosmos-physical-ai-platform,
claim-cosmos-3-omnimodal,
claim-vjepa-2-predictive-representation
|
deepmind-genie-3,
nvidia-cosmos,
nvidia-cosmos-3,
meta-vjepa-2
|
| Why autonomous driving needs world models | Traditional simulation versus learned world models, statistical mileage limits, closed-loop evidence and validation boundaries. |
claim-av-rare-scenario-pressure,
claim-av-learned-world-model-added-value,
claim-gaia-2-driving-domain,
claim-waabi-world-neural-simulation,
claim-cosmos-physical-ai-platform,
claim-av-world-models-not-road-test-replacement
|
rand-driving-to-safety,
koopman-wagner-av-safety,
wayve-gaia-2,
waabi-world,
nvidia-cosmos,
nvidia-cosmos-3
|
| How do you evaluate a world model? | Nine evaluation dimensions, route-specific metric priorities and evidence-layer boundaries. |
claim-world-model-evaluation-multidimensional,
claim-world-model-evaluation-route-specific,
claim-world-model-evaluation-closed-loop,
claim-world-model-evaluation-deployment-boundary,
claim-world-model-evaluation-evaluator-is-audited
|
world-models-2018,
deepmind-genie-3,
nvidia-cosmos-3,
meta-vjepa-2,
wayve-gaia-2,
rand-driving-to-safety,
koopman-wagner-av-safety,
mirabench-2026,
decision-centric-world-model-evaluation-2026
|
| What have world models actually achieved in robotics? | Physical-robot learning, image-goal planning, target-robot adaptation, benchmark limits and reproduction gaps. |
claim-robotics-world-model-real-hardware,
claim-vjepa-2-ac-two-lab-planning,
claim-robonet-target-robot-adaptation,
claim-robotics-world-model-protocol-sensitivity,
claim-robotics-independent-replication-gap
|
daydreamer-physical-robots,
vjepa-2-paper,
droid-dataset-v2,
robonet-multi-robot,
visual-foresight-robot-control,
dino-wm-v2,
td-mpc2-iclr,
jepa-wm-planning-tmlr,
stable-worldmodel-v2
|
| What do spatial 3D models actually measure? | Measured geometry outputs, benchmark protocol and authorship, an independent photogrammetry comparison, 3D visual reasoning and generated 3D worlds. |
claim-spatial-geometry-measured-per-output,
claim-spatialbench-protocol-bound-not-third-party,
claim-spatial-accuracy-ranking-configuration-dependent,
claim-spatial-reasoning-separate-evaluation-layer,
claim-marble-spatial-world-generation,
claim-spatial-capability-boundaries-not-established
|
dust3r-paper,
mast3r-paper,
vggt-paper,
spatialbench-v2,
aerial-3d-reconstruction-eval,
spa3r-paper,
world-labs-marble,
world-labs-about
|
Team / project source map / 团队与项目来源映射
The homepage team map is not a separate narrative layer. Its source and claim relationships are preserved here and in structured data without exposing internal IDs on the reader-facing cards.
首页团队图谱不再是独立叙述层。来源与判断关系保存在本页技术映射和结构化数据中，不再把内部 ID 直接显示在读者卡片上。
| Team / project | Route | Source ID | Claim ID |
|---|---|---|---|
| Google DeepMind / Genie 3 | Interactive generated worlds | deepmind-genie-3 |
claim-genie-3-interactive-worlds |
| NVIDIA / Cosmos | Physical-AI platform | nvidia-cosmos |
claim-cosmos-physical-ai-platform |
| Meta AI / V-JEPA 2 | Predictive representation learning | meta-vjepa-2 |
claim-vjepa-2-predictive-representation |
| OpenAI / Sora | Video generation toward world simulators | openai-sora-world-simulators |
claim-video-generation-world-simulators |
| World Labs / Marble | Spatial 3D world generation | world-labs-marble |
claim-marble-spatial-world-generation |
| Wayve / GAIA-2 | Autonomous-driving world models | wayve-gaia-2 |
claim-gaia-2-driving-domain |
| Waabi / Waabi World | Neural simulation and validation | waabi-world |
claim-waabi-world-neural-simulation |
| Runway / General World Models | Creative video toward controllable simulation | runway-gwm |
claim-runway-gwm-company-framing |
Claim source map / 判断来源映射
This table lists the main public claims rather than every sentence on the site. The full records live in claims.json.
这里列出的是主要公开判断，不是网站每一句话。完整字段、边界和复核日期保存在 claims.json。
| Claim ID | Source IDs | Status | Boundary |
|---|---|---|---|
claim-world-model-dynamics |
world-models-2018 |
unmaintained / high | Modern usage is broader than the 2018 project. |
claim-video-generation-world-simulators |
openai-sora-world-simulators |
unmaintained / high | Official framing, not proof of full interactivity. |
claim-sora-not-full-public-interactive-world-model |
openai-sora-world-simulators |
unmaintained / medium-high | Boundary judgment based on missing public environment-control evidence. |
claim-genie-3-interactive-worlds |
deepmind-genie-3 |
unmaintained / high | External validation of long-horizon stability remains limited. |
claim-cosmos-physical-ai-platform |
nvidia-cosmos, nvidia-cosmos-3 |
unmaintained / high | Platform value depends on developer adoption and downstream validation. |
claim-av-rare-scenario-pressure |
rand-driving-to-safety |
unmaintained / high | Statistical mileage result, not a universal deployment threshold or proof that learned world models are mandatory. |
claim-av-learned-world-model-added-value |
wayve-gaia-2, waabi-world, nvidia-cosmos-3 |
unmaintained / medium-high | Component and company evidence does not prove calibrated coverage or general safety value. |
claim-av-world-models-not-road-test-replacement |
rand-driving-to-safety, koopman-wagner-av-safety, wayve-gaia-2, waabi-world, nvidia-cosmos-3 |
unmaintained / medium-high | Independent sources support the validation boundary; current technical materials support the learned-model capability layer. |
claim-world-model-evaluation-multidimensional |
world-models-2018, deepmind-genie-3, meta-vjepa-2, wayve-gaia-2 |
unmaintained / medium-high | Editorial synthesis across unlike routes; no universal benchmark or independent reproduction is claimed. |
claim-world-model-evaluation-route-specific |
deepmind-genie-3, nvidia-cosmos-3, meta-vjepa-2, wayve-gaia-2 |
unmaintained / medium-high | Reusable taxonomy judgment, not a standardized score or ranking. |
claim-world-model-evaluation-closed-loop |
world-models-2018, wayve-gaia-2, nvidia-cosmos-3 |
unmaintained / medium-high | Action and inverse-dynamics evidence remains system- and protocol-specific. |
claim-world-model-evaluation-deployment-boundary |
rand-driving-to-safety, koopman-wagner-av-safety |
unmaintained / high | High-confidence autonomous-driving boundary; broader generalization is an editorial inference. |
claim-world-model-evaluation-evaluator-is-audited |
mirabench-2026, decision-centric-world-model-evaluation-2026 |
unmaintained / medium-high | Bounded to MiraBench as published on 28 May 2026, which discloses the prompt split and the per-model agreement itself; overall reported agreement is 87.8%. The position paper supplies the general principle and does not mention MiraBench. |
claim-vjepa-2-predictive-representation |
meta-vjepa-2 |
unmaintained / medium-high | Reported transfer remains task-dependent. |
claim-marble-spatial-world-generation |
world-labs-marble |
unmaintained / medium-high | Creation workflow claims do not prove strict physical simulation utility. |
claim-gaia-2-driving-domain |
wayve-gaia-2 |
unmaintained / medium-high | Driving-domain specificity limits general-purpose claims. |
claim-waabi-world-neural-simulation |
waabi-world |
unmaintained / medium | Company materials provide platform positioning, not detailed model validation. |
claim-runway-gwm-company-framing |
runway-gwm |
unmaintained / medium | Open technical detail and independent evaluation remain limited. |
Sources and references / 参考资料
These are the publications used by the Atlas. The table explains what each source is and which part of the site it supports. Machine IDs and the dated access notes remain in the structured data.
这里展示本站使用的资料、来源性质和用途；机器 ID 与带日期的访问说明保留在结构化数据中。
| Source | Type | Used for |
|---|---|---|
| World Models, Ha and Schmidhuber | Paper / project | Classic latent-dynamics framing for imagination, planning and control. |
| OpenAI: Video Generation Models as World Simulators | Official technical note | Sora's world-simulator framing and boundary between video generation and interactivity. |
| Google DeepMind: Genie 3 | Official blog / model page | Real-time navigable generated worlds and official Genie 3 capability claims. |
| NVIDIA Cosmos | Official product page | World foundation model platform for physical AI. |
| Cosmos 3 technical report | Technical report | Omnimodal world-model route across language, video, action and physical-AI scenarios. |
| Meta AI: V-JEPA 2 | Official research release | Predictive representation route, physical reasoning and robot-control transfer. |
|
RAND: Driving to Safety — official publication record Readable PDF mirror |
Independent research report | Statistical limits of road-mile validation and alternative testing as a supplement. |
| Koopman & Wagner: Autonomous Vehicle Safety | Independent peer-reviewed article | Testing-only boundary, ML validation data limits and end-to-end safety process. |
| Wayve GAIA-2 technical report | Technical report | Domain-specific autonomous-driving world-model case and rare-scenario article evidence. |
| World Labs: Marble | Official product post | Generative 3D world creation with an edit, expand, combine and export workflow. This page does not use the word “persistent”; the World Labs About page does. |
| Runway Research | Company research page | General World Models framing and creative-video-to-simulation route. |
| Waabi World | Company materials | Neural simulation, autonomous-trucking validation route and closed-loop article evidence. |
| DayDreamer | Physical-robot technical report | Online world-model learning on four separately trained robot-task setups. |
| V-JEPA 2 paper | Developer technical report | Action-conditioned post-training and two-lab Franka image-goal planning. |
| DROID v2 | Dataset paper | Robot-data provenance; arXiv metadata/PDF task-count conflict is recorded in technical notes. |
| RoboNet | Peer-reviewed paper | Multi-robot data and finetuning-based held-out robot adaptation. |
| Visual Foresight | Technical report | Video-prediction planning and physical robot control lineage. |
| DINO-WM | Technical report | Latent planning in controlled benchmark environments. |
| TD-MPC2 | ICLR 2024 paper | Continuous-control and multitask benchmark evidence. |
| What Drives Success in Physical Planning with JEPA-WMs? | TMLR reproduction study | V-JEPA 2-AC implementation and planner sensitivity, with original-author overlap. |
| stable-worldmodel-v1 | Reproduction preprint | DINO-WM Push-T reproduction and robustness stress test, with author overlap. |
| DUSt3R | Developer technical report | Pose-free point-map reconstruction and camera geometry; table identity verified, cells not read. |
| MASt3R | Developer technical report | Dense 3D correspondence and localization results; a previously recorded p. 12 locator was withdrawn as nonexistent. |
| VGGT | Developer technical report | Feed-forward geometry prediction; Section 5 states the fisheye, panoramic, extreme-rotation and non-rigid limits. |
| SpatialBench | Benchmark paper whose authors also enter a model | Cross-paradigm protocol and out-of-distribution findings. Its cross-model ranking is not third-party evaluation. |
| Evaluation of DUSt3R/MASt3R/VGGT on photogrammetric aerial blocks | Independent peer-reviewed evaluation | The one independent comparison in the spatial 3D article; the accuracy ranking reverses with image count and overlap, in aerial photogrammetry only. |
| Spa3R | Developer preprint | 3D visual-reasoning benchmark results; a single system cannot define the capability category. |
| MiraBench | Unrefereed benchmark preprint whose authors fine-tuned five of the twelve models it scores | Action-conditioned failure preservation, and the configuration of the automated evaluator that produces every score in it. |
| Decision-making-centric world-model evaluation | Unrefereed position preprint, no experiments of its own | Independent boundary on treating model-as-judge output as evidence of decision usefulness. Independent of MiraBench and does not cite it. |
Technical review log / 技术核验记录
Review notes / 核验记录
- 2026-08-08: MiraBench and the decision-making-centric position paper were added for the evaluation article's material update. Both are unrefereed preprints with no journal reference. MiraBench's own authors fine-tuned five of the twelve models it scores, and every score it reports is produced by its own VLM evaluator rather than by human annotation; both facts are recorded as source-role boundaries rather than treated as independent evidence.
- 2026-07-12: Nine robotics sources were added. Fixed PDFs, experiment tables and limitations were checked; reproduction studies remain explicitly separated from independent physical-robot validation.
- 2026-07-10: RAND Figure 3 (p. 7), Table 1 and conclusions (p. 10), plus Koopman & Wagner's testing discussion (p. 6) and conclusion (p. 9), were added as independent autonomous-driving validation evidence.
- 2026-07-10: GAIA-2 Section 2.2 (p. 6), Figure 7 and safety-critical scenario discussion (p. 13), and Cosmos 3 Table 18 / Figure 22 (pp. 65-66) were visually reviewed for page-level locators.
- 2026-07-10: Waabi World remained company-positioning evidence because no open technical report or independent benchmark was located in this review.
- 2026-07-09: Waabi World was added to the source and claim layer to match the homepage team map.
- 2026-07-09: OpenAI's Sora source returned 403 during direct access check; it remains listed as an official source with an access note.
- 2026-07-09: Genie 3 24 fps, 720p and several-minute consistency are kept as official claims, not independent validation.
- 2026-07-09: Cosmos, V-JEPA 2, Marble, GAIA-2 and Runway are treated as source-backed public claims with explicit boundaries.
These notes are a record of checks that were made, on the dates they were made. Since 2026-08-21 no further scheduled rechecks are performed, and nothing here is a statement about what a source says today.
核验记录的目的不是制造“已经盖章”的感觉，而是保留每次判断当时的依据与边界。 2026-08-21 起不再有计划性重查，这里的每一条都只说明"当时查到了什么"，不代表来源今天的状态。
Using this evidence programmatically / 供 Agent 与技术使用者
Machine IDs, field definitions, access notes and retrieval guidance now live in one technical entry point.
机器 ID、字段定义、访问说明与读取规则现在集中在一个技术入口中。