About the Project

AI agents built on large models can now plan, use tools, and carry out tasks over many steps. They are, however, very inefficient. They tend to spend the same amount of computation on easy and hard steps alike, reason at length when a quick decision would do, and call external tools without considering the cost. People are better at this: we usually put more effort into harder or more important problems and less into trivial ones.

This PhD will study how to build agents that manage their own computation: agents that decide when to think carefully, how much effort a problem deserves, and which resources to use. Several questions follow from this. How can an agent judge how difficult or uncertain a decision is, and use that judgement to decide how hard to work? How can it learn good ways of allocating effort across perception, reasoning, and action, and will those carry over to new tasks? And how should the cost of computation, such as time or energy, be weighed against the quality of the result? The student would be free to shape which of these to pursue.

The project draws on machine learning, decision theory, and work on bounded rationality in psychology. There is room for both theoretical and experimental work. The ideas matter for making AI cheaper and more practical to run, but the broader interest is in understanding how any agent should reason and act when its resources are limited.

This project would suit a candidate with a good grounding in machine learning and mathematics, and an interest in how learning systems make decisions. Some familiarity with reinforcement learning is an advantage but not required.