Large language models increasingly ship with explicit "thinking modes," but the opposite capability — reliably not thinking — has drawn far less scrutiny. A new paper accepted at NeurIPS 2026, titled "Thinking Inertia: LLMs Keep Thinking When Told Not To," argues that no-thinking deserves to be treated as a capability in its own right.
The authors first question how no-thinking is measured. Existing proxies — a disabled thinking mode or the absence of long reasoning traces — are unreliable, they argue, because disabled modes may still emit reasoning and long traces may contain filler rather than genuine inference. Instead, they normalize each response into a pre-answer trace and a final answer, then score it at three levels: an Empty-Thinking Rate for strict answer-only compliance, an instruction-aware Question-Pre-answer Relevance measure, and an LLM-as-judge Explicit Inference Rate for visible reasoning.
Using that framework, the team tested six prompting interventions across six LLMs on Boolean, multiple-choice, and open-ended questions. Explicit no-think controls could not reliably eliminate visible inference. The models instead displayed what the authors call "Thinking Inertia": explicit inference persists under strict controls and becomes more prevalent as the answer space opens.
Accuracy held steady on Boolean and multiple-choice tasks, but open-ended tasks exposed a trade-off between answer-only compliance and task accuracy. Rewriting the same questions across answer spaces showed that supplying candidate answers makes answer-only responses easier to produce.
The authors conclude that stopping explicit reasoning cannot be assumed from model settings or instructions alone, and that no-thinking should be evaluated systematically alongside reasoning ability. A project page accompanies the work.