Tuesday, 13 October 2026 · 8 min read
LLMs often continue explicit reasoning despite instructions to answer directly
This paper defines and measures "no-thinking" in LLMs as the absence of visible explicit inference before the final answer. Across six LLMs and tasks, explicit no-thinking controls rarely eliminate reasoning traces, especially on open-ended tasks, revealing a trade-off between compliance and accuracy.
Based on the paper Thinking Inertia: LLMs Keep Thinking When Told Not To by Dianqiao Lei, Kevin Qinghong Lin, Pan Lu, Philip Torr and 1 other
Challenges in defining and controlling LLM no-thinking behavior
Large Language Models (LLMs) have improved by explicitly generating intermediate reasoning steps before answers, known as "thinking modes." However, controlling or disabling this reasoning—"no-thinking"—is less understood and harder to measure. Existing approaches infer no-thinking from indirect proxies like disabling thinking modes, removing special reasoning tokens, or output length, but these are unreliable because models may still produce reasoning in ordinary text or short outputs may still contain inference.
Understanding no-thinking is important for efficiency and adaptivity: reasoning consumes tokens and latency, and humans often answer intuitively without explicit reasoning. Thus, explicit control over when LLMs reason visibly is desirable, but the community lacks a clear, operational definition of no-thinking behavior at the response level.
This paper addresses this gap by defining no-thinking as an observable property of the model's output, distinguishing between strict answer-only responses, relevant but non-inferential text, and explicit inference. It proposes new metrics to reliably measure these distinctions and studies how no-thinking varies across tasks, models, and prompting interventions.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short