Thanks for posting this. I had just ran across this one and wondered if I should post it myself. Came here to see if someone had. I'd gone looking for this after reading about DeepScaleR yesterday.<p>I'm excited about the prospect of using formal methods to generate, analytically (not via LLM), synthetic problem+solution sets of perfect quality, of progressive size & complexity, for training; and then using RL, with progressive scaling, as in DeepScaleR.