Backtracking When It Strays: Mitigating Dual Exposure Biases in LLM Reasoning Distillation

In three linesMOTAB, a new LLM reasoning distillation method, resolves dual exposure bias by dynamically monitoring student generation against an adaptive safety boundary and backtracking when it strays. Tested on LIMO-v2 and AceReason datasets, MOTAB achieves ~3% average performance improvement by mitigating both forward and reversed exposure biases.

Summary generated by Claude — human-verified