Editorial guide
How World Models Watch evaluates AI world models.
World Models Watch is not a generator, a marketplace, or a generic AI tool directory. The site follows public evidence about systems that create, simulate, reconstruct, or control worlds, then separates those systems by what people can actually do with them. A reader should be able to tell whether a page describes a usable product, an open research stack, a limited demo, a robotics platform, or a reported product signal that still needs stronger confirmation.
What counts as a world model signal?
A world model signal earns coverage when it changes how readers should understand generated environments, spatial memory, action-conditioned simulation, embodied AI, or many-agent worlds. Marble and Skybox AI matter because they give creators visible spatial outputs. Genie and Oasis matter because they make interaction and control easier to see. LingBot, Cosmos, HY-World, and related open stacks matter because they connect public code, model weights, benchmarks, or robotics workflows to the same broader category.
The page language stays conservative on purpose. If a source only proves a 360 panorama workflow, the page should not call it a persistent simulator. If a research preview is not a public API, the page should not present it as a ready developer platform. This boundary is the main value of the site: readers get a map of the category without having to treat every demo reel as the same kind of product.
How sources are ranked.
The strongest evidence is a primary source: an official project page, documentation page, GitHub repository, model card, paper, or product announcement. Reported sources can be useful for tracking early access or regional product movement, but they are marked as reported and should not carry the same weight as a release note from the organization that built the system. Every model dossier and release signal should leave a reader with a path back to the source, a short explanation of what changed, and a clear warning about what the source does not prove.
This approach also protects readers from the most common category mistakes. A video avatar model can be relevant to world modeling without being an explorable world. A robot policy model can be part of physical AI without being a consumer world creator. A many-agent simulation can be important for persistent behavior without proving that a visual generator has solved long-term world consistency.
How to use the site.
Start with the company map when you want stable profiles for specific systems. Use decision guides when you need to choose between product lanes, such as a 360 skybox workflow versus a generated 3D world workflow. Open release signals when you need to know what changed recently, then follow the related model or comparison page for the slower, more durable context.
The site intentionally links news back to stable pages instead of forcing every update to become a new standalone landing page. That keeps the public surface closer to a source-backed reference and further from a programmatic SEO wall. The goal is not to publish a page for every keyword variation; it is to give readers enough context to decide which demos are usable, which stacks are reproducible, and which claims need more evidence.
What gets updated.
Updates are prioritized when they change access, reproducibility, category boundaries, or real user choices. A new API page, a model weight release, a benchmark script, a documented pricing or queue limit, or a product rollout can all justify a new signal. A vague marketing post, a reposted demo clip, or an unsupported rumor usually should not. When the site does publish an update, the page should explain why the change matters and where the reader should verify the current status.
This means World Models Watch can stay useful even when the field moves fast. The homepage shows the visible front door, but the underlying editorial work is slower: classify the system, check the strongest source, compare it with adjacent systems, state the limit, and keep the source trail available for readers who need to inspect the claim themselves.