Online · 2026-10-21 · 08:00–09:00 PT
Talk Title: Allocating Computation Between Solving and Verification
Professor · Csaba Szepesvári · Google DeepMind
Abstract: Reasoning systems can spend computation generating candidate solutions, checking them, or trying again. How should this computation be allocated? I will discuss a stochastic solver–verifier model in which both solving and verification take random amounts of time, and verification can be noisy. The goal is to return a correct solution with high probability while minimizing expected computation. This connects universal restart strategies, sequential hypothesis testing, and adaptive scheduling. I will explore how far we can get without knowing the solver’s success probability, runtime distributions, or verifier accuracy, and discuss results and open questions concerning near-optimal allocation.
Bio: Csaba Szepesvári is a Canada CIFAR AI Chair, Professor of Computing Science at the University of Alberta, and Team Lead for the Foundations team at DeepMind. He is a Fellow of the Association for the Advancement of Artificial Intelligence and an IEEE Fellow. He received his PhD in 1999 from József Attila University in Szeged, Hungary, in probability and statistics. His research advances the foundations of learning-based artificial intelligence, especially reinforcement learning, bandit algorithms, and planning under uncertainty. He is the author or co-author of three books, including Bandit Algorithms, published by Cambridge University Press in 2020. He is also the co-inventor of UCT, an algorithm that helped ignite the modern development of Monte Carlo tree search and made simulation-based planning a central tool in game AI and decision making under uncertainty.
