Thinking Inertia: models keep reasoning even when told not to

Thinking Inertia (arXiv 2610.11765, submitted 8 October 2026), a paper led by Dianqiao Lei and featured on Hugging Face Daily Papers, studies what happens when LLMs are told not to think. Surface proxies such as a disabled thinking mode, missing reasoning tags, or short outputs often still leave question-specific reasoning before the final answer. The paper measures Empty-Thinking Rate (strict answer-only), Question-Pre-answer Relevance, and Explicit Inference Rate via a blinded judge, across six models and Boolean, multiple-choice, and open-ended tasks. Findings: native no-think switches do not reliably remove visible inference; DeepSeek-V4-Flash still showed explicit inference in 99.9% of open-ended responses under disabled thinking. Visible inference rises from Boolean to multiple-choice to open-ended tasks ("thinking inertia"). On open-ended tasks, strict answer-only prompts raised mean Empty-Thinking Rate from near zero to about 40% but cut mean accuracy by about 15 points across five models. Rewriting the same questions into Boolean or multiple-choice form made answer-only replies much easier. Authors' results; code and metrics are described in the paper. Sorted take: if you want cheap, direct answers, constrain the answer format (yes/no or choices) rather than only flipping a "no think" switch.