Entropy trajectory shape predicts LLM reasoning reliability: A diagnostic study of uncertainty dynamics in chain-of-thought
Xinghao Zhao
This paper shows that by sampling a few reasoning steps from an AI language model, you can predict whether its answer will be correct by looking at how confident it becomes—specifically, whether its uncertainty consistently decreases at every step. The key finding is that chains where uncertainty smoothly decreases at each step (monotone chains) are much more reliable (~69% accurate) than those with ups and downs (~47% accurate), and this shape pattern matters more than the total amount of uncertainty reduction. This method is cheap to compute (using only ~1,500 tokens per question) and outperforms more expensive checking methods.
chain-of-thought reasoninguncertainty quantificationcalibrationLLM evaluation