Distribution Corrected Offline Data Distillation for Large Language Models
For practitioners deploying smaller LLMs in resource-constrained settings, this work provides a more effective offline distillation method that reduces the gap between training and inference distributions.
The paper tackles distributional drift in offline reasoning distillation for LLMs, where student models trained on teacher-generated traces suffer from compounding errors during inference. The proposed method corrects this drift by adaptively emphasizing teacher supervision aligned with the student's on-policy distribution, achieving improved reasoning accuracy on GSM8K, MATH, MATH500, AMC, AIME, and OlympiadBench without online rollouts.
Distilling reasoning traces from strong large language models into smaller ones is a promising route to improve intelligence in resource-constrained settings. Existing approaches face a fundamental trade-off: offline distillation from teacher-generated traces provides high-quality, sample-efficient supervision but suffers from distributional drift: during training, the student model conditions on teacher-generated prefixes, whereas during inference the student autoregresses on self-generated prefixes, leading to compounding errors over long reasoning trajectories. Meanwhile, on-policy or self-distillation methods better match the student's inference-time distribution, but require costly online sampling and often produce low-quality traces in early training. We propose a principled offline reasoning distillation framework that preserves the efficiency and supervision quality of offline teacher-generated data while correcting teacher-student distribution drift. It adaptively emphasizes teacher supervision that is better aligned with the student's on-policy distribution. Evaluations on mathematical reasoning benchmarks of GSM8K, MATH, MATH500, and harder held-out competition-style tasks, including AMC, AIME, and OlympiadBench, show that our method improves reasoning accuracy over prior offline distillation algorithms and yields more stable reasoning traces while preserving instruction-following capabilities. Our work shows that lightweight, distribution-correction-aware training can substantially strengthen offline reasoning distillation without online rollouts.