Research direction
Questions that
move the work forward.
Our agenda evolves with what we learn. We pursue useful capabilities without fixing the lab to a single model family.
Capability
What can the model learn?
We investigate representations, architectures and training approaches that can turn new kinds of information into useful outputs.
Behavior
Where does it work?
We study performance beyond typical examples: ambiguity, incomplete information, unfamiliar inputs and the limits of generalization.
Application
What makes it useful?
We examine model quality alongside inference cost, latency and the work required to use an answer. The application informs the evaluation.
Our evaluation approach