AI interpretability scores can pick worse explanations, study finds

Original: Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability

Source: arxiv.org ↗