Abstraction Alignment: Comparing Model-Learned and Human-Encoded Conceptual Relationships

A model's confidence distribution is a reflection of its underlying knowledge.

Human abstractions represent the concepts and relationships we expect models to learn.

Abstraction alignment measures how much of a model's uncertainty can be explained by the human abstractions. SUBGRAPH PREFERENCE: Confidence in different regions of the abstraction. ABSTRACTION MATCH: Uncertainty reduced by a level of abstraction. CONCEPT CO-CONFUSION: Concepts the model regularly confuses.