Instead of having a LLM generate reflection tokens, what if we have a system of models generate reflection tokens?