Has anyone tried RL to rewrite prompts for reasoning models to further improve outputs? I'm assuming so, it feels pretty obvious, but if not I want to try it. If you know of any existing work here, pls lmk so I don't re-do something people have already done!

By this, I mean: - Take an already-trained, frozen reasoning model (i.e. o4-mini via API) - Add a smaller LLM that takes in a prompt, and rewrites it to improve how the frozen model performs - Update the smaller LLM's weights, keep the larger LLM frozen The hope is that the small LLM would learn to 'steer' the CoT of the frozen larger model better than a human could, increasing performance.

@corbtt reminded me of this work by @brendanh0gan... Brendan, how did it go? Seems pretty similar to what I'm thinking here.