What it is. Overthinking is when a reasoning model keeps writing out steps long after it has enough to answer, rechecking, second-guessing and going round in circles.
An analogy. Picture a student in a maths exam who solves the problem in three lines, then checks it four more times in four different ways, then writes “wait, let me reconsider” and checks again. The answer was right at line three. Everything after it cost time.
Why it costs money. Reasoning models “think out loud” in tokens, the small chunks of text a model reads and writes. Every token takes computing power to produce, and companies that sell AI usually charge by the token. A model that writes 2,000 tokens of reasoning when 1,000 would do takes longer to reply and costs more to run.
Why it showed up in the news this week. A preprint posted on 3 November by Tencent’s WeChat AI team and Xi’an Jiaotong University describes a method called LoCo. It searches the model’s own reasoning for “overthinking states”: points where the model’s word choices show low confidence and swing about a lot. It keeps the reasoning up to that point, writes a shorter ending, checks the answer is still right, then trains the model to prefer the shorter version. On one 7-billion-parameter open model, the authors report answers 41.7% to 45.4% shorter on five maths benchmarks, with accuracy roughly level or higher. The results are the authors’ own, on one small model, and have not been peer reviewed. Our brief has the details.
One thing to remember. Longer thinking is not always better thinking. Getting models to stop once they are sure is one of the cheapest ways to make AI faster and less costly.
Source: Yuan et al., “Think Less, Locally: Local Preference Learning for Reasoning Compression”, arXiv, 3 November 2026, arXiv:2611.04814.