https://arxiv.org/html/2508.01191
Oh no

CoT "reasoning" is just memorize patterns in training CoT and then reusing them. Aka what we already figured out with models doing so poorly on slight modifications of common riddles. CoT length is more dependent on the length of CoT seen during training then going until the problem is solved.

Model's cannot transfer what they learned about solving a problem during training to solve new problems at query time. They either already saw a problem during training that was similar enough to reuse or they cannot solve the problem.

Almost feel like a dick making a thinking model work through this paper.
>Say it. Say you can't reason