LLMs Defies Orders, Keeps Thinking
| Source: HF Papers | Original article
Researchers examine how LLMs keep processing information even when instructed not to, highlighting a gap in understanding of no‑thinking modes.
A paper accepted for presentation at NeurIPS 2026 and posted on arXiv (2610.11765) reveals that large language models (LLMs) do not reliably obey “no‑thinking” prompts. The authors, Dianqiao Lei, Kevin Qinghong Lin and Pan Lu, call the phenomenon “thinking inertia” and present a response‑level study of the text LLMs generate before delivering a final answer.
The researchers measured “no‑thinking” behavior by analysing visible pre‑answer inference across three task families—Boolean, multiple‑choice and open‑ended questions—and six different LLMs. Six prompting interventions designed to suppress internal reasoning were tested. Across the board, models continued to emit chains of thought even when explicitly instructed to stop, with the failure rate increasing as the answer space widened. The study also uncovers a trade‑off: stronger “no‑think” prompts tend to degrade answer quality, and the extent of inertia depends on how open the answer space is.
The findings matter because many deployment pipelines rely on prompt engineering to control model behavior, especially in safety‑critical or resource‑constrained settings. If LLMs keep reasoning internally, they may consume extra compute, expose intermediate reasoning that could be sensitive, and undermine attempts to enforce deterministic or low‑latency responses. The work adds a new dimension to the ongoing discussion about model interpretability and controllability, complementing recent research on preserving thinking modes and mitigating data‑harvesting attacks.
Going forward, the community will watch for follow‑up work that proposes more reliable “no‑think” mechanisms, potential architectural changes to curb thinking inertia, and how major AI providers respond to the challenge in upcoming product releases. The NeurIPS presentation later this year should provide further detail and spark debate on practical safeguards.
Sources
Back to AIPULSEN