Tracing Uncertainty in Language Model "Reasoning"

2026-05-08

AI Summary

This work addresses the limited understanding of the dynamic mechanisms underlying intermediate reasoning trajectories in current language models and the absence of effective methods to assess their correctness. The authors propose a novel approach that models generated intermediate token sequences as evolving states through the lens of uncertainty quantification, introducing for the first time uncertainty trajectory profiles—such as slope and linearity—to characterize the reasoning process. Their analysis reveals fundamental differences in uncertainty evolution between correct and incorrect reasoning paths. Experimental evaluation across five language models on GSM8K and ProntoQA demonstrates that features extracted from these uncertainty trajectories enable classification with an AUROC of up to 0.801 using only the first few hundred tokens, peaking at 0.807, substantially outperforming existing methods.

Abstract

Language model (LM) "reasoning", commonly described as Chain-of-Thought or test-time scaling, often improves benchmark performance, but the dynamics underlying this process remain poorly understood. We study these dynamics through the lens of uncertainty quantification by treating the "reasoning" traces, the intermediate token sequences generated by LMs, as evolving model states. We summarize each trace by an uncertainty trace profile: a small set of features describing the shape of the uncertainty signal over its trace, such as its slope and linearity. We find that across five LMs evaluated on GSM8K and ProntoQA, these profiles predict whether a trace yields a correct final answer with AUROC up to 0.807, improving markedly on recent related work. We reach AUROC 0.801 using only the first few hundred tokens of full traces, suggesting that errors can be detected early in the generation. A detailed comparison of correct and incorrect traces further reveals qualitatively distinct uncertainty profiles, with correct traces showing a steeper and less linear decline in uncertainty. Together, the results suggest that our method, grounded in decision-making under uncertainty, provides a principled lens for studying the generative process underlying LM "reasoning".

Problem

• language model reasoning
• uncertainty quantification
• Chain-of-Thought
• generative process
• intermediate token sequences

Innovation

• uncertainty quantification
• Chain-of-Thought
• reasoning traces
• language models
• early error detection