Csaba Szepesvari

Efficient Simple Regret Algorithms for Stochastic Contextual Bandits

Learning to Reason Efficiently with Discounted Reinforcement Learning

LACONIC: Length-Aware Constrained Reinforcement Learning for LLM

Frontier LLMs Still Struggle with Simple Reasoning Tasks

Beyond Least Squares: Uniform Approximation and the Hidden Cost of Misspecification

Eluder dimension: localise it!

REINFORCE Converges to Optimal Policies with Any Learning Rate