Autointerp provides us descriptions of LLMs features, but how it is evaluated varies from one setting to another. We propose FADE, a framework that enables standardized, automatic evaluation of alignment between features and autointerp descriptions across various metrics.

Building on the gained insights, we present a probe to track for knowledge provenance during inference and show where it is localized within the input prompt. Our attempt shows promising results, with >94% ROC AUC and >84% localization accuracy.

4/4

Using interpretability tools, we discover that heads important for RAG can be categorized into two: parametric heads that encode relational knowledge and in-context heads that are responsible for processing information in the prompt.

2/