Interpretability is the foundation of trustworthy and controllable large language model (LLM) deployment, yet unlocking actionable insights from black-box models remains a critical challenge. Sparse Autoencoders (SAEs) have emerged as a transformative solution decomposing dense LLM hidden states into sparse, human-understandable features that reveal the model’s internal mechanisms.

 

In this talk, I will discuss how SAEs enable genuine interpretability by isolating semantically coherent latent features, offering a clear window into which internal representations correspond to specific concepts. I will show core applications of this interpretability: targeted model unlearning that precisely removes unwanted knowledge without sacrificing general capability, controlled model steering that refines behavior through meaningful feature interventions, and extensions to emerging architectures like diffusion language models (DLMs) where SAEs unlock new interpretive and manipulative possibilities. I will also share key considerations for balancing interpretability, utility, and robustness, and outline how SAE-driven insights are turning theoretical interpretability into a practical tool for building more reliable, responsible AI systems. Join to explore how SAEs are redefining what it means to understand and control LLMs.