Measuring Reasoning Utility in LLMs via Conditional Entropy Reduction

Open in new window