Adversarial Concept Search: Predicting Compositional Errors From Feature Geometry

In three linesMethod to predict compositional failures in LLMs using representational geometry. When two concepts are encoded close together (linear interference), the model fails to compose them; when nearly orthogonal, it succeeds. Validated on programming, multihop reasoning, and multilingual recall.

Summary generated by Claude — human-verified