Henry Wu's training method teaches language models to self-verify, doubling math accuracy and boosting scientific reasoning 14-fold · Digg