Preprint discussed in the article:

https://arxiv.org/pdf/2508.01191

fyi, this is about Chain-of-thought, not <think>, is that still being used?

haven't read the paper closely enough to comment on the methods