Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty

TL;DR

This paper presents an information-theoretic framework to understand how large language models reason, highlighting the importance of externalizing uncertainty for effective reasoning and decision-making.

Contribution

It introduces a novel framework decomposing reasoning into procedural and epistemic components, emphasizing the role of uncertainty externalization in LLM reasoning.

Findings

Uncertainty externalization enhances reasoning performance.

Purely procedural reasoning can become informationally stagnant.

Externalizing uncertainty supports continued information acquisition.

Abstract

LLMs often exhibit Aha moments during reasoning, such as apparent self-correction following tokens like "Wait," yet their underlying mechanisms remain unclear. We introduce an information-theoretic framework that decomposes reasoning into procedural information and epistemic verbalization - the explicit externalization of uncertainty that supports downstream control actions. We show that purely procedural reasoning can become informationally stagnant, whereas epistemic verbalization enables continued information acquisition and is critical for achieving information sufficiency. Empirical results demonstrate that strong reasoning performance is driven by uncertainty externalization rather than specific surface tokens. Our framework unifies prior findings on Aha moments and post-training experiments, and offers insights for future reasoning model design.