Back to feed
arXiv cs.CL
arXiv cs.CL
6/16/2026
CoRA: Confidence-Rationale Alignment for Reliable Chain-of-Thought Reasoning

CoRA: Confidence-Rationale Alignment for Reliable Chain-of-Thought Reasoning

Short summary

CoRA addresses a critical LLM reliability issue: models can express high confidence in answers even when their reasoning is incomplete or poorly supported. Using GRPO-based reinforcement learning with rubric-based evaluation, the method jointly optimizes answer correctness and rationale quality across grounding, coherence, and task alignment. Results show up to 26.51% reduction in confidence-rationale misalignment while maintaining competitive accuracy.

  • •Framework aligns LLM answer confidence with quality of reasoning rationales
  • •Reduces confidence-rationale alignment error by 26.51% using GRPO reinforcement learning
  • •Tested on medical, math, and general knowledge tasks with improved calibration

Generated with AI, which can make mistakes.

#research-breakthrough
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Explore more

AIPO: : Learning to Reason from Active Interaction
arXiv cs.CL
Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
arXiv CS.AI
Speaking in Self-Assessing Tongues: On the Verbalized Confidence of LLMs in Machine Translation
arXiv cs.CL
Quantifying Aleatoric Uncertainty of In-Context Learning for Robust Measure of LLM Prediction Confidence
arXiv cs.CL