A Better Teacher From the Same Weights
TL;DR for operators A team running its own reasoning model has three familiar ways to improve accuracy: collect more labeled solutions, obtain a stronger teacher, or allow more computation per answer. LoOPSD1 investigates a fourth possibility: the model’s existing weights may provide better supervision when temporarily executed differently, and the improvement may survive after normal execution is restored. ...