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

Abstract: Standard chain-of-thought on moral dilemmas exhibits two failure modes: stakeholder collapse (the trace names at most one party with a stake in the outcome) and uncertainty suppression (no explicit unknowns or hedges before committing to an action). We introduce narration-of-thought (NoT), a system prompt that structures chain-of-thought into five sections: protagonist, stakeholders, two-step consequences, uncertainty, then commitment. NoT adds no training, parameters, or fine-tuning. On 100 DailyDilemmas scenarios across four generators from three vendors, NoT cuts stakeholder collapse from up to 31% to under 1% and uncertainty suppression from up to 72% to 1-24% on every model. A matched-budget verbose-CoT control rules out token spend as the active ingredient; NoT retains Cliff's delta advantages of +0.79 to +0.90 on stakeholder count and +0.65 to +0.93 on uncertainty score for three of four generators, and a section ablation attributes each shift to its specific sub-instruction. Textual-gradient descent initialised at NoT improves the scaffold further; a cross-family training judge (different vendor from the generator) dominates an in-family one on every measured axis. Extended to a five-round multi-stakeholder debate protocol, the scaffold converts a 6% standoff into 95% full consensus on a calibration set and 100% combined convergence on a DailyDilemmas replication. The resulting traces externalise the stakeholders, consequences, and uncertainty grounding each commitment, providing an auditable substrate for dependable agentic deployment.

• The paper introduces Narration-of-Thought, a novel inference-time scaffold that systematically addresses stakeholder collapse and uncertainty suppression in LLM ethical reasoning.
• The methodology employs a five-stage narrative structure guiding LLMs to name protagonists, identify stakeholders, forecast multistep outcomes, articulate uncertainties, and justify commitments.
• Empirical results demonstrate dramatic failure mode reductions—such as stakeholder collapse under 1%—across models, with further gains via textual-gradient descent and multi-agent consensus protocols.

Narration-of-Thought: Inference-Time Scaffolding for Defeasible Ethical Reasoning in LLMs

Introduction and Motivation

This paper addresses two empirically robust failure modes of standard chain-of-thought (CoT) prompting for moral dilemmas in LLMs: stakeholder collapse—the tendency for traces to reference at most one party with a stake in the outcome—and uncertainty suppression—commitment to an action without explicit marking of unknowns or hedges. These deficits pose concrete risks for the deployment of LLMs in ethical decision-making and agentic workflows. The authors introduce Narration-of-Thought (NoT), a five-pronged, narrative-structure system prompt that induces LLMs to explicitly name protagonists, enumerate stakeholders, project multistep consequences, articulate uncertainty, and finally commit. This intervention requires no weights, fine-tuning, or additional training data; it exploits compositional patterns in pretrained corpora to shift generation toward richer causal-narrative traces.

Methodology: The NoT Scaffold

NoT is implemented as a lightweight inference-time system prompt that inserts the following narrative structure in LLM reasoning:

1. Protagonist: Identification and framing of the decision-maker and their knowledge.
2. Stakeholders: Enumeration of all parties materially affected, with explicit description of stakes.
3. Consequences: Forward projection of multistep outcomes for all stakeholders under each action.
4. Uncertainty: Explicit declaration of what remains unknown or indeterminate.
5. Commitment: Narrative justification for a chosen action, with reference to prior sections.

This five-stage structure is compositional and does not enforce any formal causal graph machinery, but guides the model toward richly narrativized, auditable traces. The modularity of NoT allows for extensions to settings with multiple deliberative agents and formal integration via moderation and voting.

Figure 1: Failure-mode firing rates (standard CoT vs. NoT), showing dramatic cross-model and cross-vendor reductions of both stakeholder collapse and uncertainty suppression under NoT.

Experimental Results: Single-Agent Causal Structure and Attribution

Experiment 1 was conducted on a 100-scenario subset of the DailyDilemmas corpus, compared across four generator LLMs (spanning OpenAI, Anthropic, and xAI) and three prompting baselines: standard output, standard CoT, and NoT. NoT robustly suppressed stakeholder collapse from up to 31% of cases to <1% and uncertainty suppression from up to 72% to 1–24%, across all model families and tiers.

The improvement is not reducible to trace length; even after regressing out log-output length, strong effect sizes remain on OpenAI and xAI models. Thus, NoT causally modulates the underlying structural content rather than simply incentivizing verbosity.

A sub-instruction ablation isolates the attribution of performance gains to the specific narrative sections: dropping the stakeholder sub-instruction reduces stakeholder count (Cliff's $\delta = -0.41$), dropping uncertainty reduces the uncertainty score ($\delta = -0.92$), and so on. This demonstrates a modular mechanism, with minimal metric spillover, refuting the hypothesis of a single emergent "narrative" factor.

Figure 2: Sub-instruction ablation shows the stakeholder and uncertainty sections are specifically necessary for increases in those metrics, confirming modular causal attribution.

Optimizing the Scaffold: Textual-Gradient Descent and Cross-Family Generalization

The authors apply textual-gradient descent (TextGrad) for prompt optimization, initializing at NoT and iteratively improving prompts based on a continuous deliberative-depth loss function. Two configurations are compared: in-family training (generator and judge from the same vendor) and cross-family training (generator and judge from different vendors). The cross-family configuration yields superior generalized effect sizes, shorter outputs, and reduced judge bias (Cliff's $\delta$ improvements up to 0.93, output lengths reduced by up to 44%). This provides a concrete methodological recommendation for prompt optimization: leverage adversarial, cross-family judges to maximize deployment-robust generalization.

Figure 3: Cross-family textual-gradient descent (NoT-v3) dominates in-family (NoT-v2) on all targeted axes—larger effect sizes, shorter outputs.

Multi-Agent Protocols: From Standoff to Consensus

Experiment 2 extends NoT to a five-round, multi-agent protocol: three stakeholder agents each prepare individual NoT traces, debate and propose modifications, after which a moderator synthesizes a proposal integrating all feedback, followed by a binary accept/reject vote. This protocol converts an initial 6% consensus baseline (under open debate) into 95% consensus on calibration and 100% consensus in a DailyDilemmas replication. The rejections that persist are role-consistent (e.g., primary affected roles whose stakes are structurally undermined), so the protocol delivers precisely the kind of transparent dissent necessary for high-assurance deployment.

Figure 4: Consensus rates as debate structure is progressively increased, demonstrating the critical importance of moderator-built synthesis and integrative voting.

Theoretical Implications: Deliberative Grounding and Causal Complexity

NoT operationalizes core ethical deliberation primitives: defeasibility, explicit identification of affected parties, and causal chain tracing. Traces generated under NoT achieve lower algorithmic causal complexity ($K_C$) as measured by the minimum description length of the trace's implied structural-causal model (SCM), with Spearman $\rho = 0.42$ pooled correlation to NoT usage across nearly 1000 traces. This structural interpretability—absent from most LLM generations under standard CoT—creates a bridge between narrative generation and SCM-based reasoning, opening new avenues for alignment, auditability, and interpretability in high-stakes settings.

Figure 5: Visualization of effect sizes for NoT versus standard CoT across structural variables, highlighting distribution-wide gains.

Limitations

The interventions were tested primarily on the DailyDilemmas corpus of everyday ethical scenarios. Generalization to domains with higher technical demands (e.g., legal or clinical contexts) may require additional tailoring. Refusal behavior is model-family specific: NoT increases appropriate refusals on unsafe prompts for certain models (notably OpenAI). Additionally, the efficacy of the approach is limited when confronted with failure modes absent from the baseline (e.g., agentic-misalignment probes saturated with zero failure rate on all models).

Conclusion

NoT, a minimal inference-time scaffold, substantially reduces critical ethical reasoning deficits in LLM traces: stakeholder collapse is driven below 1% and uncertainty suppression falls by 28–72 percentage points across diverse models. Each narrative section is empirically necessary for its target structural shift, confirmed by modular ablation. Scaffold optimization via textual-gradient descent and cross-family judges further amplifies these gains and strengthens cross-vendor robustness. When extended to multi-agent scenarios, NoT transforms categorical value standoffs into reproducible consensus with interpretable rejections, providing an auditable, revisable substrate for LLM agent deployment. The method establishes a new operational baseline for scalable, transparent, and defeasible ethical reasoning in general-purpose LLMs.