Best of N I mean. Run inference multiple times, then pick the best result. This would explain how you can enable/disable advanced mode in the same chat without changing the model.