Emanuele Marconato

Faithful concept bottleneck explanations are impossible without including all concepts, but probabilistic modeling and group lasso enable compact, faithful explanations.

Relational knowledge in LLMs isn't as simple as a fixed linear mapping between subject and object embeddings, and how you ask the question matters.

KL divergence can fool you: matching a teacher's predictions doesn't guarantee preserving its internal representations, but distilling with logit distance does.