Intelligence needs a useful understanding of the world beyond the immediate prompt. We are interested in how systems represent their environment, learn its structure, and update their understanding as new evidence arrives.
Questions we’re interested in
- How can a system learn representations that generalize across unfamiliar environments?
- How should a world model distinguish observation, inference and uncertainty?
- How can understanding of cause and effect support reliable action?
Our starting point
Our direction connects learned representations, causal reasoning, memory and interaction. Simulation is one tool for exploring these questions, alongside the broader challenge of learning from and acting in the world.
What this means today
A convincing representation is not proof of real-world understanding. Generalization, predictive accuracy and robustness need independent evaluation. We have not published results establishing those capabilities.
Reasoning & planning