CS Colloquium

Time and Location:

Speaker:

Link:

Traditional AI systems rely on models that execute a fixed set of computations at test time. But many of the hardest problems we care about require test-time adaptation: models must perform variable computation at test time, adapting what they do to the specific instance in order to make progress on hard, open-ended tasks. For example, when given a conjecture in research mathematics, a model may need to break the problem into lemmas, test intermediate claims, revise failed proof attempts, and allocate more computation to promising directions. To succeed in such settings, we need to train models to behave as effective algorithms. This is now possible because pre-trained models come with powerful priors that make test-time adaptation feasible. In this talk, I introduce a framework that casts algorithm learning as meta–reinforcement learning, providing a principled foundation for both formal analysis and the practical design of training objectives and methods for learning algorithms. I conclude by showing how these methods can train a 4B theorem-proving model that acts as an effective algorithm and outperforms models up to 30x larger.