CoRA aligns model confidence with rationale quality in chain-of-thought reasoning

A new arXiv paper introduces CoRA, or Confidence-Rationale Alignment, for improving chain-of-thought reasoning. The method trains models so confidence in an answer is better supported by the generated rationale.

The problem is that models can sound confident while providing rationales that are incomplete, weakly grounded, or only plausibly connected to the answer. That mismatch can make explanations misleading even when the final answer is right.

CoRA uses reinforcement learning with rewards for correctness, answer probability, and rubric-based rationale support, targeting more reliable reasoning behavior.