Narration-of-Thought: Inference-Time Scaffolding for Defeasible Ethical Reasoning in Large Language Models

Patrick Cooper and colleagues introduce narration-of-thought (NoT), a system prompt requiring no training or fine-tuning, which structures chain-of-thought reasoning into five sections. Across 100 DailyDilemmas scenarios and four model generators from three vendors, NoT reduced stakeholder collapse from up to 31% to under 1% and uncertainty suppression from up to 72% to 1-24%. A verbose control ruled out token spend as the cause, and section ablations confirmed each sub-instruction's specific contribution.