1/8 🧵 Did you know that it’s possible to steer LLM behavior at each token on the fly? Excited to share my recent #ICLR2025 workshop paper in collaboration with afterless, nev, and Sheikh Abdur Raheem Ali (many, many thanks😆) Paper: arxiv.org/abs/2505.03189 Poster (TLDR):

🧵1/6 SAEs have become a staple of LLM interpretability, but what if we applied them to image generation models? My recent paper with Dmitrii Kharlapenko, Yixiong Hao, afterless, Sheikh Abdur Raheem Ali, and Arthur Conmy adapts SAEs to understand the SOTA diffusion transformer FLUX.1 ⬇️