Mechanistic Anomaly Detection
Definition
Using a causal explanation of a model's training-time behavior to flag deployment inputs where that behavior occurs for a different internal reason — detecting, for example, a deceptive model whose 'am I being trained?' check has flipped.
Cast within
No topic-region cast yet — this appears once Mechanistic Anomaly Detection's compiled claims are aligned into a topic region's argument tree.
Claims mentioning Mechanistic Anomaly Detection (0)
No claims found for this concept.
My Notes
Loading notes...