Validity-Calibrated Reasoning Distillation for LLMs

A new framework for reasoning distillation, called validity-calibrated reasoning distillation, is proposed in arXiv:2605.04078. Unlike traditional methods that treat distillation as trajectory imitation with static teacher-student hierarchies, this approach frames it as local learning-signal allocation. It compares the student's and teacher's next-step actions under the same prefix and uses their relative local validity to modulate the distillation update, addressing the misalignment where intermediate steps are locally under-specified. The method yields dynamic, context-dependent updates.