Shuaiyi Nie, Dingsiyu, Wenyuan Zhang, Linhao Yu, Tianmeng Yang, Yao Chen, Weichong Yin, Yu Sun, Hua Wu, Tingwen Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The significant computational demands of large language models have increased interest in distilling reasoning abilities into smaller models via Chain-of-Thought (CoT) distillation.Current CoT distillation methods mainly focus on transferring teacher-generated rationales for complex reasoning to student models.However, they do not adequately explore teachers' dynamic attention toward critical information during reasoning.We find that language models exhibit progressive attention shifts towards key information during reasoning, which implies essential clues for drawing conclusions.Building on this observation and analysis, we introduce a novel CoT distillation framework that transfers the teacher's stepwise attention on key information to the student model.This establishes structured guidance for the student's progressive concentration on key information during reasoning.More importantly, we develop a Mixture of Layers module enabling dynamic alignment that adapts to different layers between the teacher and student.Our method achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets.To our knowledge, it is the first method to leverage stepwise attention within CoT distillation to improve small model reasoning. Input Question(a) A sample from the SVAMP dataset.The distilled student model fails to adequately utilize numerical information, leading to erroneous results, whereas the teacher model, during stepwise reasoning, effectively utilizes all numerical information to arrive at the correct final result.