[論文レビュー] Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty

Jeonghye Kim, Xufang Luo+4 more|arXiv (Cornell University)|Mar 16, 2026

Logic, Reasoning, and Knowledge被引用数 0

ひとこと要約

論文は推論を手続き情報と認識的言語化に分割する情報理論的枠組みを提案し、不確実性を外在化することが情報獲得の継続と強い推論性能の鍵であると示す。表面トークンの「Wait」などにとどまらない。

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.

研究の動機と目的

• Motivate a formal account of reasoning in LLMs beyond pure procedural step-by-step execution.
• Introduce epistemic verbalization as externalized uncertainty that guides subsequent reasoning.
• Define information sufficiency and analyze how externalized uncertainty aids information gain when procedural reasoning stalls.
• Demonstrate empirically that uncertainty externalization, not surface tokens, drives strong reasoning and self-correction.

提案手法

• Formalize reasoning as self-Bayesian inference with augmented states that separate procedural and epistemic components.
• Define information gain and information sufficiency to measure progress toward the correct answer.
• Characterize limits of purely procedural reasoning and identify failure modes under divergence.
• Introduce epistemic verbalization as an externalizable epistemic signal that enables continued information acquisition.
• Experiment with test-time manipulation of epistemic tokens and distillation to study the role of epistemic verbalization in learning and performance.

実験結果

リサーチクエスチョン

• RQ1 What informational roles do procedural reasoning and epistemic verbalization play in LLM reasoning under uncertainty?
• RQ2 How does externalizing uncertainty affect information gain and information sufficiency during reasoning trajectories?
• RQ3 Do tokens like “Wait” causally reflect epistemic verbalization, or are they surface indicators of a deeper mechanism?
• RQ4 How does distillation and training that preserve epistemic verbalization impact reasoning performance?

主な発見

• Epistemic verbalization enables continued information acquisition even when procedural reasoning stalls, driving information sufficiency.
• Token-level uncertainty (e.g., entropy of next-token) does not reliably predict progress toward the correct answer; trajectory-level epistemic signals are crucial.
• Epistemic verbalization, not specific tokens, correlates with stronger reasoning and self-correcting behavior across models and tasks.
• Suppression of epistemic tokens degrades performance, while inducing them or using few-shot prompts with epistemic cues improves reasoning performance.
• Distillation that preserves epistemic verbalization is essential for effective transfer; removing epistemic uncertainty signals harms performance, even with correct procedural traces.