Validity-Calibrated Reasoning Distillation

# Validity-Calibrated Reasoning Distillation Researchers propose a reasoning distillation framework that replaces token-level trajectory imitation with validity-calibrated learning signals, where student model updates are weighted by the relative correctness of student vs. teacher next-step actions under identical prefixes, addressing the under-specification problem in multi-step reasoning transfer.