a paper with a similar idea on scaling test time reasoning, this is sorta how all the thinking models work under the hood. https://arxiv.org/abs/2501.19393