Embodied Graph world model plans over graph states in Minecraft and Crafter
A new arXiv paper proposes the Relational World Model, which imagines future outcomes over compact graph states rather than dense pixel observations, and reports diamond results in Minecraft.
In plain words
AI agents that act in games or on the web need to plan far ahead, but predicting every pixel is wasteful and error-prone. The paper's Relational World Model instead compresses each situation into a small graph of task-relevant entities and imagines how that graph evolves. In the authors' Minecraft tests, it was the only tested method to obtain diamonds within the budget.
What was announced
The authors argue that existing world models, which learn latent dynamics for imagination-based planning, rely on dense pixel-level observations. This forces the model to absorb redundant information while overlooking relations among task-relevant entities, so small errors can accumulate over long imagined rollouts and corrupt predictions, they write.
Their answer is Embodied Graph, a relational state abstraction, paired with the Relational World Model that imagines directly over graph states. It combines a constructor that filters redundant details and maintains graph memory through online updates with a recurrent state space model that learns node-wise latent dynamics while capturing dependencies between nodes. The imagined graph states predict action consequences to guide decisions. In experiments on Minecraft and Crafter, RWM was the only method to obtain diamonds within the given budget, and web interaction tests suggested relational state prediction can inform web agent decisions.
What it does not show
- The evidence is the paper's abstract only; experimental details, baselines and budgets are not described, and the reported results are the authors' own and are not independently confirmed.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.