Why try this idea on base models only? The whole point of reasoning models is to automatically use COT and related techniques to bring out more capabilities.

Why try this idea on base models only? The whole point of reasoning models is to automatically use COT and related techniques to bring out more capabilities. It would be interesting to see if this is doing anything that’s not already being exploited.