The Problem: Ethical Reasoning Failure Modes
Standard chain-of-thought (CoT) prompting often fails in moral dilemmas due to two primary issues:
- Stakeholder Collapse: The model fails to identify multiple parties affected by a decision, often focusing on only one.
- Uncertainty Suppression: The model commits to a decision without acknowledging potential unknowns or hedging, leading to overconfident and potentially harmful outputs.
The Solution: Narration-of-Thought (NoT)
Narration-of-Thought is a system prompt that structures the reasoning process into five mandatory sections:
- Protagonist: Defining the actor.
- Stakeholders: Explicitly listing all affected parties.
- Two-step consequences: Mapping out potential outcomes.
- Uncertainty: Identifying what is unknown or ambiguous.
- Commitment: Making the final decision based on the previous steps.
This approach requires no fine-tuning or additional parameters, relying entirely on inference-time scaffolding. By forcing the model to externalize these components, the reasoning becomes auditable and more robust.
Impact and Performance
Testing across 100 scenarios with four different model generators showed significant improvements:
- Stakeholder Collapse: Reduced from up to 31% to under 1%.
- Uncertainty Suppression: Reduced from up to 72% to between 1% and 24%.
- Debate Protocol: When extended to a five-round multi-stakeholder debate, the scaffold converted a 6% standoff rate into 95% full consensus on calibration sets, demonstrating that the structure helps models converge on more nuanced, agreed-upon ethical positions.
The authors confirmed that these gains are due to the structural instructions rather than simply increasing token count, as a matched-budget verbose-CoT control group did not achieve similar results.