Research interests
I’m interested in how AI uses learned relationships in new situations and predicts the consequences of its actions.
Generalization
Which relations must a model preserve to reason in a new situation?
I’m particularly interested in how models represent relationships and changes in their environment. Can a model still use these relationships to predict and plan when the objects or conditions change?
Possible approaches
Vary the learned representation and prediction objective, then test unseen combinations of relations and changes in the environment.
What to evaluate
Transfer, predictive accuracy, and the quality of downstream decisions.
Planning depends on representations that preserve how actions change a situation. Testing transfer can help reveal which of those relationships a model has learned.
Read the related research ↗Predicting action outcomes
Can an agent recognize when its next move would close off its goal?
In scientific search, an agent may edit a composition or a sequence over several steps. I’m interested in whether it can predict how each edit affects the chance of reaching the goal.
Possible approaches
Begin with environments where goal reachability can be checked, then introduce partial observations and changing constraints.
What to evaluate
Goal attainment, preservation of feasible paths, constraint violations, and computational cost.
For planning, a world model needs to predict how an action affects later choices. Constraints such as a limited edit budget make it possible to test whether those predictions are useful.
Read the related research ↗Information gathering
When should an agent observe, consult another agent, or ask a person?
An agent may need to check a source or consult someone before acting. I’m interested in how it decides what information to seek, and whether that information changes its decision.
Possible approaches
Control what is visible and remembered; compare additional observations, peer evidence, and expert input under a shared cost budget.
What to evaluate
Decision improvement, information cost, correction of errors, and whether people can understand and act on requests.
Active learning asks which observations are worth acquiring. In collaborative tasks, a related question is who to ask and how to use the answer.
Read the related research ↗Applications
Scientific search is one setting for these questions: an agent proposes a change, evaluates it, and decides what to try next. I’m also interested in tasks where an agent needs to consult a person before proceeding.
Email me