AI RESEARCH
Beyond the Black Box: Identifiable Interpretation and Control in Generative Models via Causal Minimality
arXiv CS.LG
•
ArXi:2512.10720v2 Announce Type: replace Deep generative models, while revolutionizing fields like image and text generation, largely operate as opaque ``black boxes'', hindering human understanding, control, and alignment. While methods like sparse autoencoders (SAEs) show remarkable empirical success, they often lack theoretical guarantees, risking subjective insights. Our primary objective is to establish a principled foundation for interpretable generative models.