HEAL：基于回顾熵辅助学习的推理蒸馏

A novel framework called Hindsight Entropy-Assisted Learning (HEAL) is proposed to address the limitations encountered when distilling reasoning capabilities from Large Reasoning Models (LRMs) into smaller models. Traditional methods treat the teacher as a static filter, neglecting the exploration of complex 'corner-case' problems where the teacher fails to independently identify valid solutions, thus artificially creating a 'Teacher Ceiling' for the student. HEAL synergizes three core modules, drawing from the educational theory of the Zone of Proximal Development (ZPD): Guided Entropy-Assisted Repair (GEAR), an active intervention mechanism that detects critical reasoning breakpoints to enhance the learning effectiveness of the student model. This approach is RL-free and aims to bridge the reasoning capability gap, providing new insights and methods for reasoning distillation.