Mechanistic-Interpretability

A new ICML paper shows language models detect uncertainty internally — occupying representation regions with 2-3× the intrinsic dimensionality of factual inputs — but the signal never reaches the output. Hallucination isn’t ignorance. It’s a severed connection between knowing and speaking.

A new paper reveals why prompt injection keeps working despite safety training: models assign authority based on how text sounds, not where it comes from. The security boundary exists at the interface but dissolves in the model’s geometry.