AI Seminar: Actionable Interpretability for Language Model Reasoning
This seminar presents research on actionable interpretability, translating insights into language model reasoning, learning, and failure modes into concrete improvements in training and inference. Topics include parameter-efficient adaptation via localized computation, distillation dynamics, and gradient-based analysis for reward hacking suppression.