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

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Narration-of-Thought (NoT) is a novel system prompt designed to enhance Large Language Model (LLM) ethical reasoning by structuring their chain-of-thought. It addresses two common failure modes in standard CoT on moral dilemmas: stakeholder collapse and uncertainty suppression. NoT constrains the model's output to five narrative sections: naming a protagonist, enumerating stakeholders, projecting two-step consequences, articulating uncertainty, and then committing to a decision. This protocol requires no training data, parameters, or fine-tuning. On a 100-scenario DailyDilemmas sample, NoT reduced stakeholder collapse from up to 31% to under 1% and uncertainty suppression from up to 72% to 1–24% across four generators from three vendors. A matched-budget verbose-CoT control confirmed the narrative scaffold, not just increased token spend, was the active ingredient. The method also extends to a multi-stakeholder protocol, converting a 6% debate standoff into 95% full consensus.

Key takeaway

For AI Scientists or Machine Learning Engineers developing LLM-powered agents, implementing Narration-of-Thought (NoT) system prompts can significantly improve ethical reasoning. This approach forces explicit consideration of stakeholders and uncertainties, reducing critical failure modes like stakeholder collapse and uncertainty suppression. The resulting auditable decision traces enhance dependability for agentic workflows. Consider using cross-family judges for prompt optimization to ensure robustness across different LLM vendors.

Key insights

Narration-of-Thought (NoT) prompts improve LLM ethical reasoning by structuring output into narrative primitives, reducing critical failure modes.

Principles

• Scaffold LLM reasoning into domain-specific primitives.
• Cross-family training judges improve prompt optimization.
• Narrative structure grounds causal reasoning in LLM outputs.

Method

NoT is a system prompt instructing LLMs to generate a five-section trace: protagonist, stakeholders, two-step consequences, uncertainty, and commitment, before a final decision.

In practice

• Implement NoT prompt for ethical dilemma analysis.
• Use cross-vendor judges for prompt optimization.
• Structure multi-agent debates with moderator integration.

Topics

• Large Language Models
• Ethical AI
• Prompt Engineering
• Chain-of-Thought
• Multi-Agent Systems
• Defeasible Reasoning

Code references

• PatrickAllenCooper/ANI_Computational_Narratology

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, AI Ethicist, Machine Learning Engineer

Related on AIssential

• Narration-of-Thought: Inference-Time Scaffolding for Defeasible Ethical Reasoning in Large Language Models — Narration-of-Thought (NoT) enhances LLM ethical reasoning by structuring chain-of-thought to explicitly address stakeholders and uncertainties.
• Code Reasoning for Software Engineering Tasks: A Survey and A Call to Action — LLM code reasoning improves through structured CoT, execution feedback, and inference scaling, culminating in agentic systems.
• Breaking the Chain: A Causal Analysis of LLM Faithfulness to Intermediate Structures — LLMs' apparent faithfulness to intermediate reasoning structures is fragile; they often fail to update predictions after causal intervention.
• Why Sampling Is Not Choosing: Intentionality, Agency, and Moral Responsibility in Large Language Models — LLMs lack intrinsic intentionality and self-attributed action, precluding moral responsibility despite normative outputs.
• Staying with the Uncertainty: Uncertainty-Scaffolding Strategies for Artificial Moral Advisors in LLM-to-LLM Simulated Conversations — Uncertainty-scaffolding strategies enhance engagement quality in LLM-based Artificial Moral Advisors during ethical discussions.
• Teaching Machines to Code: A Social Scientist’s Guide to LLM-Based Information Extraction — Multi-agent LLM pipelines excel at complex information extraction by mimicking human collaborative coding processes.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.