This research introduces a novel betting game framework that trains large language models to accurately express confidence in their answers, enhancing safety and trustworthiness.

Key Innovations:

• Uses reinforcement learning to calibrate LLM confidence estimations
• Models confidence as a betting game with rewards for accurate self-assessment
• Penalizes both over-confidence and under-confidence
• Specifically targets factual question-answering scenarios

For security applications, this approach significantly reduces the risk of LLMs providing misleadingly confident but incorrect information, which is crucial for deployment in high-stakes environments like healthcare or critical infrastructure.

Rewarding Doubt: A Reinforcement Learning Approach to Confidence Calibration of Large Language Models