[2603.18940] Entropy trajectory shape predicts LLM reasoning reliability: A diagnostic study of uncertainty dynamics in chain-of-thought

Computer Science > Computation and Language arXiv:2603.18940 (cs) [Submitted on 19 Mar 2026 (v1), last revised 27 Mar 2026 (this version, v2)] Title:Entropy trajectory shape predicts LLM reasoning reliability: A diagnostic study of uncertainty dynamics in chain-of-thought Authors:Xinghao Zhao View a PDF of the paper titled Entropy trajectory shape predicts LLM reasoning reliability: A diagnostic study of uncertainty dynamics in chain-of-thought, by Xinghao Zhao View PDF HTML (experimental) 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. Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG) Cite as: arXiv:2603.18940 [cs.CL]   (or arXiv:2603.18940v2 [cs.CL] for this version)   https://doi.org/10.48550/arXiv.2603.18940 Focus to learn more arXiv-issued DOI via DataCite Submission history From: Xinghao Zhao [view email] [v1] Thu, 19 Mar 2026 1...