[UltraSane]

This paper [1] even claims that "models primed with incorrect solutions containing proper reasoning patterns achieve comparable performance to those trained on correct solutions."

[1] https://arxiv.org/abs/2503.01307