Transfer Learning for Weak-to-Strong Generalization in Language Model Alignment

Watch a 42-minute lecture from the Simons Institute where Yuekai Sun from the University of Michigan explores the possibility of aligning large language models (LLMs) with superhuman capabilities using human feedback without compromising their performance. Learn how weak-to-strong generalization can be achieved through transfer learning by eliciting latent knowledge from pre-trained LLMs. Discover why traditional fine-tuning approaches face fundamental limitations and explore an alternative refinement-based method that effectively addresses these challenges. Examine practical applications through three LLM alignment tasks that demonstrate the effectiveness of the refinement approach in domain adaptation and related areas.