Distribution Corrected Offline Data Distillation for Large Language Models

Addresses distributional drift in offline LLM distillation by correcting the mismatch between teacher-conditioned training and student self-generated inference to reduce compounding errors in long reasoning traces.