Inverse Reasoning Models for Efficient Post-Training and Verification from Answer-Only Data

A self-distillation technique using answer-only data trains an inverse reasoning LLM to generate chain-of-thought traces, enabling more efficient post-training without requiring supervised reasoning data.