Sources

Framing & taxonomy

• id: li-taxonomy

  Fei-Fei Li & World Labs, "A Functional Taxonomy of World Models," Substack, June 3, 2026. https://drfeifei.substack.com/p/a-functional-taxonomy-of-world-models — The renderer / simulator / planner taxonomy; POMDP framing; "statistical structure of space and time" contrast with LLMs; the drone-shot renderer critique. Republished excerpt: https://www.a16z.news/p/a-functional-taxonomy-of-world-models

• id: barenholtz-counterpoint

  Elan Barenholtz, Substack note responding to the taxonomy, 2026. https://substack.com/@generativebrain/note/c-271034560 — The sequence-prediction counterargument: continuing signals may buy the behavior (syntax, melody, "physics") without an explicit inner model. Use as the strongest form of the opposing view in Level 4.

Foundational papers (stable, canonical)

• id: ha-schmidhuber-2018

  Ha & Schmidhuber, "World Models," 2018. https://worldmodels.github.io/ (interactive paper) / arXiv:1803.10122 — The VAE + RNN + controller recipe; training an agent inside its own dream.

• id: sutton-dyna

  Sutton, "Dyna, an Integrated Architecture for Learning, Planning, and Reacting," SIGART Bulletin, 1991. — Practice on an internal model; the ancestor of learning-in-imagination. URL pending verification

• id: dreamerv3

  Hafner et al., "Mastering Diverse Domains through World Models" (DreamerV3), arXiv:2301.04104. — The Minecraft-diamond result; the most legible demonstration of learning inside a latent dynamics model.

• id: muzero

  Schrittwieser et al., "Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model" (MuZero), Nature, 2020. — Model only what matters for value, policy, and reward. URL pending verification

2025–2026 systems

• id: genie3

  Google DeepMind, "Genie 3: A new frontier for world models," DeepMind blog, 2025. — Real-time interactive world generation, ~24fps at 720p, consistency over minutes; closed API as of 2026. URL pending verification

• id: qwen-agentworld

  Qwen Team, "Qwen-AgentWorld: Language World Models for General Agents," June 24, 2026. Blog: https://qwen.ai/blog?id=qwen-agentworld — arXiv: https://arxiv.org/abs/2606.24597 — GitHub: https://github.com/QwenLM/Qwen-AgentWorld — HF: https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B — Native language world model simulating 7 agent domains (MCP, Search, Terminal, SWE, Web, OS, Android); predicts next environment observation; used as a decoupled simulator for agentic RL; AgentWorldBench.

• id: dreamzero

  NVIDIA GEAR, "DreamZero: World Action Models are Zero-shot Policies," 2026. Project: https://dreamzero0.github.io/ — arXiv: https://arxiv.org/abs/2602.15922 — World Action Model on a video diffusion backbone; jointly predicts future world states and actions; real-time closed-loop control (~150ms/action chunk); 2x generalization over VLA baselines in reported experiments.

• id: cosmos

  NVIDIA, "Cosmos World Foundation Model Platform for Physical AI," arXiv:2501.03575. — Open, self-hostable world foundation model platform for robotics/AV; the report itself acknowledges remaining object-impermanence and gravity violations.

• id: marble

  World Labs, Marble (launched commercially November 2025). https://www.worldlabs.ai/ — Persistent 3D scene generation from text/images; exports Gaussian splats and meshes; the spatial-intelligence camp. URL pending verification

• id: vjepa2

  Meta AI, "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning," arXiv:2506.09985 + Meta AI blog overview. — Prediction in representation space; V-JEPA 2-AC action-conditioned variant for robot planning. URL pending verification

• id: dreamer4

  Hafner et al., "Training Agents Inside of Scalable World Models" (Dreamer 4), 2025. — The current generation of the Dreamer line; agents optimizing behavior inside a scalable learned simulator. URL pending verification

• id: gaia

  Wayve, GAIA-2 / GAIA-3 — generative world models for driving, positioned as offline evaluation infrastructure. URL pending verification

Surveys & synthesis (for orientation and the "camps" framing)

• id: zero-to-hero

  "World Models, From Zero to Hero," HackMD, 2026. https://hackmd.io/@AbdelStark/world-model-from-zero-to-hero — The five-camps table: video generation / spatial-3D / generative WM / latent JEPA / infrastructure; action-conditioned as the key column.

• id: kenhuang-architectures

  Ken Huang, "World Models, Architectures, and the Next Phase of AI," Substack, May 2026. https://kenhuangus.substack.com/p/world-models-architectures-and-the — The three-camp framing (agent-centric, foundation-WM, spatial-intelligence); encode/roll-forward/extract decomposition; architecture families.

• id: graison-2026

  Graison Thomas, "Beyond the Video Hype: Why World Models Feel Different in 2026," Medium, April 2026. https://medium.com/@graison/beyond-the-video-hype-why-world-models-feel-different-in-2026-88486a295fe3 — Closed-loop evaluation becoming central (World-in-World, WorldArena); platformization; open problems list (long-horizon coherence, memory, physical realism, benchmark comparability, sim-to-real).

• id: mts-drop

  MTS, "World Models," drops.mts.now, May 2026. https://drops.mts.now/world-model/ — General explainer; mentions DreamZero and Agora-1. NOTE: Agora-1 could not be independently verified in open sources as of 2026-07-02 — do NOT include Agora-1 on the site unless a primary source is added here.

Context (funding, scale — use sparingly, secondary sources)

• id: introl-race

  "World Models Race 2026," Introl blog, Jan 2026. https://introl.com/blog/world-models-race-agi-2026 — LeCun's AMI Labs (€500M at €3B valuation), Genie 3 release, Marble commercial launch, Cosmos 2M downloads. Secondary source; prefer primary sources for any claim that survives to prose.

Curation rules

1. Landscape claims on the site trace to these ids. No system, date, or capability claim may come from the builder's training knowledge alone.
2. URL pending verification entries: confirm the link before citing in visible prose; until then the claim may appear only with the source listed as "title, venue."
3. When updating the site, add sources here first, then edit `/content/landscape.json`, then bump the Last-updated date.