Fixing language model training by correcting wrong reasoning paths

Original: Trajectory-Refined Distillation

Source: arxiv.org ↗