GRPO Framework Cuts LLM Confidence-Rationale Misalignment by 26.5%

A GRPO-based RL framework jointly optimizes answer correctness, committed-answer probability, and rubric-graded rationale quality without exposing gold answers to the judge. Tested on MedQA, MathQA, and OpenBookQA across three open-weight LLMs, it reduces confidence-rationale alignment error by up to 26.51% versus SFT and correctness-only GRPO baselines while maintaining accuracy.