"Foundations of Post-training from Base Models" – Audrey Huang, Talks at TTIC скачать в хорошем качестве

"Foundations of Post-training from Base Models" – Audrey Huang, Talks at TTIC

“Foundations of Post-training from Base Models” Audrey Huang, University of Illinois Urbana-Champaign Originally recorded on January 5, 2026, at TTIC. In this talk, Audrey Huang examines the theoretical foundations of post-training in large language models and how it interacts with pre-trained base models to elicit complex reasoning behavior. She introduces key quantities that characterize base-model coverage over high-quality responses and develops optimal algorithms within this framework, offering new insights into how reinforcement learning can extend model capabilities beyond pre-training. Timestamps: 00:00 Introduction 01:30 Talk begins 54:25 Q&A #LLMs #ReinforcementLearning #PostTraining #MachineLearningTheory #AI #LearningTheory #Optimization #Research #TTIC