Rethinking Circuit Evaluation: Do Circuits Explain Model Errors?

arXiv:2609.35686v1 Announce Type: cross Abstract: Mechanistic interpretability (MI) aims to explain a model's behaviour through analyzing its internal computations; circuit-based explanations aim to isolate these computations with compact subnetworks validated