Models Trained on Self-Explanations Show Generalization

Adam Karvonen described training models on thousands of explanations of their own behaviors such as ignoring requests or coding errors. The single general dataset produced generalization to held-out evaluations. John Schulman called the work timely given declining CoT monitorability and noted that a metric for explanation quality enables hill-climbing via counterfactual simulatability. Laura Ruis linked related work from her group. Owain Evans amplified the thread. The posts present the training results and positive reactions without further external confirmation.