Great observation. Maybe an additional “routing model” could be trained to predict when it’s better to think more vs just using the current result