Entropy trajectory shape predicts LLM reasoning reliability: A diagnostic study of uncertainty dynamics in chain-of-thought

TL;DR

This study introduces a diagnostic method based on the shape of entropy trajectories to assess and improve the reliability of reasoning in large language models, especially in uncertain scenarios.

Contribution

It presents a simple, interpretable, and model-agnostic approach to understanding uncertainty dynamics in LLM reasoning, enhancing selective prediction and triage.

Findings

Entropy trajectory shape effectively predicts reasoning reliability.

Method is practical, inexpensive, and robust across models and datasets.

Provides insights into uncertainty dynamics in numeric and discrete-answer tasks.

Abstract

Understanding uncertainty in chain-of-thought reasoning is critical for reliable deployment of large language models. In this work, we propose a simple yet effective diagnostic approach based on trajectory shape rather than scalar magnitude. We show that this signal is practical, interpretable, and inexpensive to obtain in black-box settings, while remaining robust across models and datasets. Through extensive ablations and cross-domain replications, we demonstrate its utility for selective prediction and triage. Our findings offer a generalizable insight into uncertainty dynamics in reasoning tasks, with particular focus on numeric and discrete-answer settings.