For much of the past two years, one of the clearest ideas in frontier AI has been that reasoning models get better when they are given more time to think.
The recipe sounds intuitive. Instead of asking a model to jump straight to an answer, let it generate a long chain of intermediate reasoning. Give it more test-time compute, and difficult problems become easier. In effect, the model gets to spend more time working things out before committing to an answer.
But research emerging in 2026 is putting an important qualification on that idea: sometimes the model has already solved the problem, and then talks itself out of the correct answer.
This phenomenon is increasingly being called “overthinking”.
Why an LLM Overthinks
A Google DeepMind study published in July examined the thought processes of open-source reasoning models including Qwen3 and distilled versions of DeepSeek-R1.
Rather than simply counting how many tokens a model uses, the researchers broke its reasoning into smaller sub-thoughts and mapped how those thoughts connect. They identified patterns they call the “Explorer” and the “Late Landing”. Both point to over-exploration and over-verification as important drivers of unnecessary reasoning.