Can AutoResearch train its model on the best results of its optimization runs?

A user asks whether anyone is researching AutoResearch with model training in the loop: using reinforcement learning on the best results of optimization runs instead of a static model like Fable or GPT. They ask how well it works and for resources. A reply suggests borrowing ideas from a linked post to get past the initial phase before fine-tuning, if starting with a minimally chat-aligned model.