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

When AI models face moral dilemmas, they often ignore the people affected. They may also act too confidently without acknowledging unknowns. Current reasoning methods tend to flatten complex social situations into thin, decisive judgments. This research introduces a new prompting method called Narration-of-Thought (NoT). It forces the AI to tell a story about the people involved and the uncertainties before making a decision. This leads to much more thorough and balanced reasoning.

The failure of step-by-step reasoning

As Large Language Models (LLMs) enter sensitive workflows, engineers face new challenges. Standard "chain-of-thought" (CoT) prompting—simply asking a model to "think step by step"—is insufficient for ethical reasoning. While CoT helps with math, it fails to capture the situational structure of a moral crisis.

The authors identify two repeatable failure modes in standard CoT traces. The first is "stakeholder collapse." This occurs when the model's reasoning mentions at most one party affected by the decision. This effectively ignores the broader social impact. The second is "uncertainty suppression." Here, the model commits to an action without naming explicit unknowns or using linguistic hedges (words like "perhaps" that signal doubt).

As shown in, a standard CoT trace might recognize a single constraint.

However, it may fail to realize that a decision ripples out to affect commuters, local businesses, and emergency responders. These deficits create deployment risks. An agent that ignores stakeholders or suppresses uncertainty is prone to sycophancy (agreeing with the user to please them) and misalignment.

Structuring thought through narrative

Instead of relying on abstract logic, the authors propose Narration-of-Thought (NoT). This is a system prompt that constrains reasoning into five specific narrative sections. This approach moves the model toward the dense narrative patterns found in its training data. Think of it like moving from a dry spreadsheet to a detailed case study.

The NoT protocol follows a strict sequence: 1. Protagonist: The model describes the decision-maker and their knowledge. 2. Stakeholders: The model lists all parties whose lives intersect the decision. 3. Consequences: The model projects outcomes at least two steps forward for every stakeholder. 4. Uncertainty: The model explicitly states what remains unknown about those futures. 5. Commitment: The model finally commits to a decision within the narrative frame.

By enforcing this structure, the researchers aim to externalize causal links. This creates an auditable substrate (a readable trail of evidence). Humans can inspect this trail to ensure decisions are grounded in reality.

Evidence of increased reasoning depth

The authors report that NoT significantly improves reasoning quality across multiple frontier models. Testing on a 100-scenario sample from the DailyDilemmas corpus, the study finds that NoT cuts stakeholder collapse from as high as 31% to under 1%. It also reduces uncertainty suppression from 50–72% down to between 1% and 24% .

For a developer, this means the model is far more likely to account for diverse human interests.

Crucially, the paper shows these improvements are not just caused by "verbosity." Verbosity is the tendency for longer prompts to yield more information. To test this, the authors used a matched-budget control. They compared NoT against a very verbose standard CoT prompt with the same token limit. Even with the same amount of "room" to speak, NoT outperformed the control in stakeholder count and uncertainty scoring [Table 2].

The researchers also explored automatic optimization. Using "textual-gradient descent"—a process where an LLM diagnoses and rewrites a prompt—the authors found a key trick. They discovered that training the optimizer with a "cross-family" judge (a model from a different vendor than the generator) yielded superior results. This approach produced prompts that were more effective and more concise .

Scaling from individuals to debates

The research extends this scaffold into a multi-stakeholder protocol. This setup is designed to resolve conflicts. In this version, three distinct agents represent different perspectives. They engage in a five-round debate. A moderator then synthesizes their positions into a single proposal for a final vote.

The authors report that this protocol converts a 6% debate standoff rate into a 95% full consensus rate on calibration scenarios .

Most importantly, the protocol preserves "defeasibility." This is the ability to revise a conclusion when new information or objections arise.

When agents reject a proposal, the authors find that those rejections are "role-appropriate." They concentrate among stakeholders whose interests are fundamentally undermined. This prevents the model from simply complying with a majority view. Instead, it allows developers to see exactly why a consensus failed.

Limits of the narrative scaffold

The authors note several limitations. The effectiveness of NoT was tested on the DailyDilemmas corpus. This focus is on everyday personal and civic quandaries. It is still unknown if this scaffolding transfers to technical domains like clinical triage or legal analysis. Those fields may require more than simple storytelling.

Furthermore, the intervention affects models differently. For the gpt-5.4-nano model, NoT increased both appropriate refusals and "over-refusal" (declining to answer safe prompts) [Table 8]. This suggests that added deliberative depth can make a model more cautious. Developers may need to calibrate this based on their specific use case.

Finally, while the scaffold improves reasoning, it is not a total safety solution. The authors emphasize that NoT is an auditability and interpretability tool. It is not a replacement for fundamental alignment.

The verdict: an auditable substrate for agents

For engineers building agentic workflows, the verdict is promising. If you need autonomous agents to be accountable, NoT offers a low-cost mechanism. It ensures those agents "show their work" in a human-readable way.

This method requires no fine-tuning or new parameters. It is very cheap to implement during inference (the phase where the model generates a response). The research shows that by changing how we ask a model to think, we can transform shallow responses into structured deliberations. Code for the implementation is reportedly available; see the paper for the canonical link.

Figures from the paper