Intel / article

Early Stopping in Large Reasoning Models: A Survey and Empirical Audit of Efficiency and Reliability

Source: arxiv.org
Author: Quang Hung Nguyen, Hanh Nguyen, Thanh Tam Nguyen, Quoc Viet Hung Nguyen, Trinh Pham, Thanh Trung Huynh
Date: Published Nov 3, 2026

Context

Reasoning models often keep thinking after they already have enough to answer, and early-stopping methods try to cut that reasoning short to save . This survey and audit asks whether the savings hold up once you count the cost of deciding when to stop and check that the answers stay reliable (arXiv). The authors compare methods across five reasoning models on math, science, code generation and factuality tasks, replaying the same recorded reasoning runs for every method so the comparison is fair. They also propose a measure, Target-Integrated Token Efficiency (TITE), which scores a method's token use against an oracle that stops at the ideal point, across a range of accuracy targets.

The abstract reports reliability problems rather than a winning method. Answers taken at later points in the reasoning were sometimes less correct than the authors' model of random decoding noise predicts, so thinking longer did not always help. Some models kept reasoning after being forced to stop, and some early-stopped answers that were correct came with weaker supporting reasoning. The abstract does not give method-by-method numbers, so which techniques fare best is left to the full paper.

Key takeaways · AI-distilled

The audit compares early-stopping methods across five reasoning models on math, science, code generation and factuality tasks, using a shared rollout replay protocol.
The authors argue must account for the cost of deciding when to stop, and introduce Target-Integrated Token Efficiency (TITE), which measures efficiency relative to an oracle across accuracy targets.
Beyond continued reasoning after forced termination, the reliability checks found weaker supporting reasoning in some correct early-stopped outputs, so a right answer did not always come with a sound justification.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source arxiv.org