ObserverBench Evaluates Mechanistic Interpretability for Closed-Loop Control

Accuracy in mechanistic interpretability does not guarantee effective interventions. ObserverBench tests whether internal estimators can successfully direct activation steering or circuit removal, showing that error propagation in closed-loop control requires evaluating action-loss rather than just estimation accuracy.