Thinking Inertia: LLMs Keep Thinking When Told Not To
“No-thinking” instructions did not reliably stop models from showing inference.
The paper tests six prompting interventions across six LLMs and finds that explicit reasoning can persist even under strict controls. The effect is stronger on open-ended questions than on Boolean or multiple-choice tasks. Supplying candidate answers made answer-only responses easier to produce. Accuracy stayed stable on constrained tasks, while open-ended tasks showed a trade-off between answer-only compliance and task accuracy. HF Daily Papers' note
The paper tests six prompting interventions across six LLMs and finds that explicit reasoning can persist even under strict controls. The effect is stronger on open-ended questions than on Boolean or multiple-choice tasks. Supplying candidate answers made answer-only responses easier to produce. Accuracy stayed stable on constrained tasks, while open-ended tasks showed a trade-off between answer-only compliance and task accuracy. HF Daily Papers' note
score 5