Work

Confidence Leaps in LLM Reasoning: Early Stopping and Cross-Model Transfer

Year: 2026

Type: conference-paper

Abstract: We challenge the assumption that large language models (LLMs) build confidence gradually while reasoning.Instead, we find that confidence often appears in a discrete "moment of insight", a sharp increase in the probability of the correct answer, which we call a confidence leap.Leveraging this finding, we introduce a training-free, model-agnostic early-stopping heuristic.When a confidence leap is present, we can halt the generation, reducing reasoning tokens by up to 31% without accuracy loss.We further show that the reasoning preceding a leap is semantically transferable: providing this partial reasoning to a different model family substantially improves its accuracy.These results suggest that confidence leaps mark shared, interpretable reasoning milestones rather than model-specific artifacts.

Source: Conference of the European Chapter of the Association for Computational Linguistics

DOI: 10.18653/v1/2026.eacl-short.43

Authors: Pavel Tikhonov, Ivan Oseledets, Elena Tutubalina

Institutions: Skolkovo Institute of Science and Technology, AIRI - Artificial Intelligence Research Institute

Language: English

Topic: Topic Modeling

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funder: National Research University Higher School of Economics

OpenAlex ID: W7140127370