Early Stopping for Large Reasoning Models via Confidence Dynamics

Parsa Hosseini, Sumit Nawathe, Mahdi Salmani, Meisam Razaviyayn Soheil Feizi · 2026

Read Early Stopping for Large Reasoning Models via Confidence Dynamics by Parsa Hosseini, Sumit Nawathe, Mahdi Salmani, Meisam Razaviyayn Soheil Feizi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-04930.txt.

Description

Hosseini and four colleagues tackle overthinking in large reasoning models, where long chains of thought cost compute and can even worsen answers. They track confidence in intermediate answers and find a pattern: correct paths often reach high-confidence answers early, while wrong paths ramble with shaky confidence. Their CoDE-Stop method uses those confidence dynamics to halt reasoning at the right moment. It needs no extra training and slots easily into existing pipelines, aiming to save compute while keeping answers sound.

Catalog record

CREATORParsa Hosseini, Sumit Nawathe, Mahdi Salmani, Meisam Razaviyayn Soheil Feizi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41288
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.04930

Browse the full library · About the Tower of Babel