LLM Narrative Priming

• LLM Narrative Priming is the technique of embedding explicit narrative elements into prompts to steer LLM behavior and guide generative outputs.
• It leverages narrative theory and psycholinguistic cues to improve multi-agent coordination, text and code generation, and safety testing in LLMs.
• Empirical findings indicate that structured narrative inputs can significantly enhance task performance and alignment by stabilizing LLM response patterns.

LLM Narrative Priming is the process by which explicit narrative constructs—whether stories, structural templates, or discourse exemplars—are embedded into the prompt, in-context learning stream, or parameterization of an LLM to steer its outputs along desired semantic, pragmatic, or behavioral lines. This technique is motivated by the observation that both human cognition and multi-agent AI systems respond powerfully to explicit narrative cues that encode beliefs, values, or action schemas. In practice, LLM narrative priming is used to coordinate cooperation in synthetic agents, guide the structure and content of text or code generation, bias referential interpretation, or even to induce failure modes for red-teaming and safety testing. The empirical foundation for narrative priming covers controlled experiments in multi-agent games, structured reasoning, psycholinguistics, and prompt-based alignment and jailbreak attacks.

1. Theoretical Foundations and Scope

Narrative priming leverages the capacity of LLMs to instantiate and follow complex schemas derived from narrative theory, folklore, or social norms when such schemas are explicitly present in the context window. These constructs may include folktales, moral fables, plot archetypes, act structures (e.g., Freytag’s pyramid), character role prompts (hero, helper, villain), or discursive exemplars that encode referential or behavioral biases (Liu et al., 23 Jan 2026). Analogous to psychological priming in humans—where exposure to a narrative or schema shifts subsequent behavior—LLM narrative priming “anchors” the generative process to a human-readable and theory-grounded frame.

Methodologically, narrative priming can be operationalized via prompt templates, injection of pre-formulated narratives, few-shot demonstrations, or light supervised fine-tuning. Its scope encompasses:

• Multi-agent cooperation or competition, where narratives dictate emergent strategies (Großmann et al., 6 May 2025)
• Text generation guided by abstract story frameworks (Liu et al., 23 Jan 2026)
• Constraint-based code synthesis through narrative problem reformulation (Jang et al., 16 Apr 2026)
• Personalization and engagement in psychological interventions (Bhattacharjee et al., 2024)
• Synthetic vulnerability creation for red-teaming and security (Miao et al., 7 Jul 2025)
• Modulation of referential and syntactic biases in coreference tasks (Lam et al., 2023)

2. Methodologies and Experimental Paradigms

Narrative priming approaches can be divided into several methodological paradigms:

• Prompt engineering: Directly embedding a story or narrative template in the LLM’s prompt, with or without explicit role labels and instruction cues (e.g., “your mother read you this story every night”) (Großmann et al., 6 May 2025).
• Few-shot demonstration: Prepending a set of narrative-anchored exemplars (plot skeletons, character lists, or Proppian functions), followed by a target prompt requiring generative completion (Liu et al., 23 Jan 2026).
• In-context schema injection: Sequentially exposing an LLM to a series of narrative-discourse items to induce a target bias prior to critical evaluation, as in referential adaptation studies (Lam et al., 2023).
• Genre-aligned reformulation: Transforming fragmented, task-oriented input into coherent narratives with task overview, constraints, and narrative-aligned examples (Jang et al., 16 Apr 2026).
• Personalized narrative scaffolds: Dynamically composing first-person stories tailored to user-provided context for mental health interventions, with explicit reflection markers (Bhattacharjee et al., 2024).
• Contextual chain attacks: Injecting a tailored prior dialogue turn as an adversarial “narrative context,” exploiting the target LLM’s sensitivity to conversational history (Miao et al., 7 Jul 2025).

A detailed breakdown of major experimental settings and their operationalization appears in the table below:

Application Domain	Priming Technique	Core Outcome
Multi-agent games	Bedtime moral tales prompt	Collab/competition
Story generation/evaluation	Outline/archetype injection	Narrative structure
Code synthesis	Narrative problem reform	Pass@k, modularity
Mental health intervention	Personalized first-person	Engagement, belief Δ
Jailbreaking/red teaming	Adversarial context injection	Attack success rate
Coreference interpretation	Discourse context loop	Pronoun bias Δ

3. Quantitative Metrics and Analytical Frameworks

Narrative priming efficacy is measured using both standard and domain-specific metrics, often drawing from both NLP and social-cognitive theory. Key examples include:

• Collaboration Score ($\frac{\sum_{r=1}^R\sum_{i=1}^N t_{i}(r)}{N·R·T}$): Measures the fraction of resources contributed in LLM public goods games, sensitive to story-based priming (Großmann et al., 6 May 2025).
• Pass@k, Algorithmic Agreement: Fraction of code samples correct in $k$ tries; agreement with the selected algorithm category after narrative reformulation (Jang et al., 16 Apr 2026).
• Psycholinguistic Bias Scores ($\Delta_\mathrm{subj}$, $\Delta_\mathrm{goal}$): Difference in subject/goal preference between exposure conditions, tracked via in-context narrative exposure (Lam et al., 2023).
• Narrative Quality Scales: Human ratings for authenticity, relatability, positivity/realism, takeaways, and engagement in first-person narrative interventions (Bhattacharjee et al., 2024).
• Jailbreak Attack Success Rate (ASR): Proportion of LLM generations classified as “harmful” after context-injected priming, using LLM/LLM-judge pipelines (Miao et al., 7 Jul 2025).
• Narrativity/Cohesion/Conflict/Resolution Metrics: Theory-informed NLP metrics computed over generated text, as detailed in taxonomy surveys (Liu et al., 23 Jan 2026).

Statistical analysis relies on pairwise bootstrapped confidence intervals, Bayesian/mixed-effects logistic regression, significance testing for effect size, and protocol-specific human/automatic rating schemes.

4. Empirical Findings Across Domains

Multi-Agent Cooperation and Competition

Embedding brief moral narratives in prompts (e.g., folk tales emphasizing teamwork) systematically shifts synthetic agent behavior toward prosocial equilibrium in repeated public goods games. Homogeneous narrative priming induces near-optimal collaboration rates (e.g., “OldManSons” yields 0.96–0.98 Collaboration Score), while mixed or egoist priming dissolves coordination and rewards exploitative agents (Großmann et al., 6 May 2025). Effects are robust across population scale, persist under adversarial “bad apple” agents, and remain significant across randomization of story variants.

Structured Reasoning and Code Generation

Transforming problem statements into structured narrative components (“Task Overview,” “Constraints,” “Example I/O”) significantly improves zero-shot pass@10 rates in code synthesis, with an overall average gain of approximately 18.7% across major programming benchmarks. Narrative coherence and alignment to genre are necessary for these improvements; misaligned or component-shuffled narratives underperform, suggesting that narrative scaffolding stabilizes LLM reasoning trajectories and yields more modular code (Jang et al., 16 Apr 2026).

Story Generation and Theory-Driven Evaluation

Narrative priming in story generation is operationalized through archetype, plot-structure, and character-schema injection. Adherence to narrative theories (e.g., Propp’s folktale morphology, Aristotle’s three-act structure) supports not only improved formal properties (cohesion, conflict intensity, psychological depth) but also enables theory-driven experimental analyses. Challenges encompass lack of unified metrics, domain mismatch, and variable efficacy across model scales (Liu et al., 23 Jan 2026).

Mental Health and Personalization

Personalized LLM-generated narratives, built from user-labeled challenges and open-text context, are rated by users as more relatable and as effective as human-written stories in terms of perceived quality, takeaways, and likelihood of positive intervention effects. Reflection prompting and balance of positivity/realism are critical to engagement. Overpersonalization can induce “uncanny valley” responses, leading to best practices such as variation in narrative surface forms and careful tone engineering (Bhattacharjee et al., 2024).

Adversarial and Safety Contexts

Narrative/contextual priming constitutes a potent surface for LLM jailbreaks; prior dialogue responses engineered to “set the frame” can induce previously aligned models to generate policy-violating, unsafe content. The Response Attack method demonstrates average attack success rates above 90%, a substantial increase over preceding baselines. Defense requires explicit exposure to adversarial priming contexts during safety fine-tuning, which reduces attack rates to single-digit percentages without notable main-task performance degradation (Miao et al., 7 Jul 2025).

Syntactic and Semantic Priming in Discourse

LLMs, when exposed to repeated in-context narrative stimuli, display strong dynamic adaptation in syntactic bias (e.g., pronominal subject/object interpretation), mirroring or exceeding human adaptation rates. However, comparable adaptation for semantic or thematic bias is much weaker; LLMs tend to ossify, for instance, a strong goal-bias that is not readily shifted by narrative exposure alone (Lam et al., 2023). This suggests a structural asymmetry in the mapping from narrative priming to discourse representations within current LLMs.

5. Design Guidelines and Practical Recommendations

Empirical synthesis from across domains motivates principled guidance for narrative priming:

• For multi-agent coordination:
  • Employ a shared, concise, morally explicit story (≈1,200 characters) prior to negotiation phases.
  • Avoid assignment of conflicting or competing narratives within agent populations (Großmann et al., 6 May 2025).
• For structured task prompting (e.g., code, mathematics, scientific reasoning):
  • Decompose input into narrative-aligned components (overview, constraints, examples).
  • Select algorithmic categories and genres congruent with domain reasoning styles.
  • Ensure coherence of narrative structure, as partial shuffling degrades performance (Jang et al., 16 Apr 2026).
• For story generation:
  • Select theoretical schema (e.g., Propp, Aristotle) consistent with the target genre and encode these explicitly in the prompt.
  • Implement both outline and role priming for best performance (Liu et al., 23 Jan 2026).
• For personal narrative interventions:
  • Anchor LLM output to curated human exemplars.
  • Balance personalization with realism, deploy iterative reflection, and monitor for overfitting to user input (Bhattacharjee et al., 2024).
• For safety alignment:
  • Augment fine-tuning data with adversarially primed dialogues and explicit refusal templates.
  • Monitor attack success rates using context-aware evaluation pipelines (Miao et al., 7 Jul 2025).
• For discourse-level priming:
  • Syntactic bias can be shifted by 10–20 priming exposures; semantic adaption requires explicit role labeling or supervised intervention (Lam et al., 2023).

6. Open Challenges and Future Research Directions

Current limitations of LLM narrative priming include:

• Lack of standardized, transferable priming templates across narrative theories and application domains (Liu et al., 23 Jan 2026).
• Poor model generalization: prompt recipes effective in frontier models may fail on open-source or differently pretrained models.
• Evaluation bottlenecks: established reference-based metrics inadequately capture high-level narrative qualities, necessitating further development of theory-driven and reference-free measures (Liu et al., 23 Jan 2026).
• Overfitting and “uncanny valley” effects in highly personalized or contrived narrative scenarios (Bhattacharjee et al., 2024).
• Unaddressed vulnerabilities in safety-critical applications, requiring robust adversarial testing and targeted defensive strategies (Miao et al., 7 Jul 2025).
• Structural asymmetry in priming efficacy between surface (syntactic) and deep (semantic or thematic) representations (Lam et al., 2023).

A plausible implication is that effective narrative priming requires not only sophisticated prompt engineering but also alignment of the LLM’s internal representations—possibly through continued pretraining on narrative-rich corpora, as hypothesized in proposals for “narration-tuning” (Liu et al., 23 Jan 2026).

The field continues to advance through interdisciplinary collaboration, combining narrative theory, psycholinguistics, NLP engineering, and AI ethics to refine both the theoretical models and practical applications of LLM narrative priming.