World models
How do models learn the structure of an environment and anticipate what happens next? We examine where predictions stay reliable and where changes in the environment break them.
Research directions.
Four connected questions shape the lab’s work, from learning a world to acting within it.
Interpretability
What information is encoded inside a model, and which signals actually drive a prediction? We compare observed representations with controlled interventions on model behavior.
Embedded world models
How can an agent use an internal model to compare possible actions before acting? This direction studies the relationship between imagined outcomes, uncertainty, and choice.
Continual learning
How can a model adapt when evidence arrives over time or its environment changes? We are interested in what is retained, what is revised, and when adaptation becomes unstable.