Narration-of-Thought Improves LLM Ethical Reasoning and Transparency.

A new system prompt, termed "narration-of-thought" (NoT), has been developed to improve the ethical reasoning capabilities of large language models (LLMs). Traditional chain-of-thought methods often fail by either ignoring multiple stakeholders or suppressing uncertainty when addressing moral dilemmas. NoT addresses these shortcomings by structuring the reasoning process into five distinct sections: identifying the protagonist, stakeholders, two-step consequences, uncertainties, and finally, the commitment. This method requires no additional training, parameters, or fine-tuning for the LLMs. Experiments conducted across 100 DailyDilemmas scenarios, using models from three different vendors, showed remarkable improvements. NoT reduced stakeholder collapse from up to 31% to under 1% and uncertainty suppression from up to 72% to a mere 1-24% across all tested models. A control experiment confirmed that these improvements were due to the structured prompt, not just increased token usage. The resulting reasoning traces externalize critical elements like stakeholders, consequences, and uncertainties, providing a transparent and auditable foundation for deploying dependable AI agents. Further refinements using textual-gradient descent and cross-family training judges enhanced the scaffold, demonstrating its potential to foster consensus in multi-stakeholder debate protocols.