Martingale Score: Evaluating Bayesian Rationality in LLM Reasoning

This ArXiv paper introduces the Martingale Score, an unsupervised metric designed to assess Bayesian rationality in Large Language Model (LLM) reasoning. The research contributes to the growing field of LLM evaluation, offering a potential tool for improved model understanding and refinement.

• •Introduces a novel unsupervised metric (Martingale Score) for evaluating LLM reasoning.
• •Focuses on assessing Bayesian rationality within LLMs.
• •Potentially aids in better understanding and improving LLM performance.