[에레혼#24] Sparse Autoencoders for Mechanistic Interpretability: Advances, Evaluation, and Limitations (김재원 연구원)