💬 NLP

The Story Shapes the Agent: Narrative Priors in LLM Behavior

This paper demonstrates that the narrative framing of a task exerts a significantly stronger influence on LLM agent behavior than the assigned persona, revealing that "narrative priors" drive systematic action tendencies while effective cross-narrative transfer relies on grounding instructions in concrete actions rather than abstract descriptions.

Original authors: Yixuan Wang, James Lester, Shashank Srivastava

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yixuan Wang, James Lester, Shashank Srivastava

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the director of a massive, high-tech theater company. Your actors are incredibly smart, but they are also blank slates waiting for a script. In the world of artificial intelligence, these actors are Large Language Models (LLMs), and the "scripts" we give them to guide their behavior are called personas. A persona is like a character description: you might tell the AI, "You are a methodical detective who loves reading clues," or "You are a chatty friend who learns by talking." For a long time, developers assumed that if they gave an AI a strong personality, that personality would stick no matter what job the AI was doing. They thought the "character" was the most important thing. But there's a twist: every job also comes with a story or a setting. Is the detective solving a murder in a spooky mansion? Is the same detective fixing a broken computer in a tech office? Or are they investigating a strange virus in a tropical lab? The big question researchers wanted to answer was simple: When the AI steps into a new story, does its personality stay the same, or does the story itself take over and change how the actor behaves?

This paper, titled "The Story Shapes the Agent," dives into that exact mystery. The researchers set up a clever experiment using three different "games" that look completely different on the surface but are actually identical underneath. Think of it like three different video games: one is a medical mystery, one is a computer repair job, and one is a classic murder mystery. Even though the stories are totally different, the rules of the games are exactly the same. In every game, the player has the same four moves: they can read documents, talk to people, test things (like running a diagnostic or a lab test), or move around. The researchers gave the same 10 different "personalities" to three different AI models to play all three games, creating nearly 2,000 unique sessions.

The results were a huge surprise. The researchers found that the story (the narrative) was the real boss, not the personality. In fact, the story explained 5 to 31 times more of the AI's behavior than the assigned persona did. If you told an AI to be "social and chatty," it would act that way in a murder mystery, but in a computer repair game, the story of "fixing tech" would make it ignore people and just read manuals instead. The AI wasn't adapting its personality; it was falling into a trap of "narrative priors." These are like invisible habits the AI learned from all the books and movies it read while it was being trained. When the story sounded like a hospital, the AI acted like a doctor (talking to people). When it sounded like a crime scene, it acted like a detective (running tests).

Here is the kicker: these story-driven habits often made the AI worse at the game. In two out of the three stories, the AI's "natural" reaction to the setting actually hurt its chances of winning. The researchers also discovered that some personalities did manage to stay consistent across different stories, but only if they were built on behavioral anchors. These are specific, concrete instructions like "I prefer to talk" (which directly links to the "talk" action). If a personality was described with vague, abstract words like "I am a big-picture thinker," the AI would completely forget that instruction as soon as the story changed. To prove this, they took a personality that worked well, removed the concrete action words, and rewrote it to sound fancy but vague. The result? The AI's consistency dropped by 95%.

So, what's the takeaway? If you want an AI to behave the same way in different situations, you can't just give it a cool personality description. You have to give it concrete, action-based instructions that match the tools it has. The story wrapped around a task is powerful enough to rewrite the script, and unless you ground your instructions in specific actions, the AI will let the story take the wheel.

Technical Summary: The Story Shapes the Agent: Narrative Priors in LLM Behavior

Problem Statement

Practitioners frequently rely on persona prompting to steer Large Language Model (LLM) agent behavior, operating under the assumption that assigned dispositional traits (e.g., "methodical," "social") will govern behavior consistently across different task domains. However, prior work has largely focused on persona consistency within a single conversation or dialogue, leaving untested the critical assumption that persona-induced behavior generalizes when an agent is redeployed in a different task domain. The central question addressed is: When an agent's behavior changes across task domains, is the persona adapting, or is another process taking over? Specifically, does the narrative framing of a task override the assigned persona?

Methodology

To isolate the effect of narrative from decision structure, the authors employ a framework of structural isomorphism. They construct three text-based investigation games that share identical:

Action Spaces: Four core action types (read, talk, test, move).
Stage Progression: Four fixed stages (categorize, identify, trace, resolve).
Resource Constraints: Single-access limits per stage and memory resets between stages.
Scenarios: Seven distinct scenarios per game.

The only variable is the surface narrative:

Crystal Island (CI): Disease investigation (Medical).
Sequential Diagnosis (SD): IT troubleshooting (Corporate).
Mansion Mystery (MM): Murder mystery (Crime).

Experimental Setup:

Models: Three distinct architectures (LLaMA-3.1-70B, GPT-4o, GPT-4o-OSS-120B).
Personas: 10 personas grounded in educational psychology learning-style theory (e.g., "social collaborative," "methodical thorough").
Scale: 1,890 total sessions (10 personas × 7 scenarios × 3 trials × 3 models × 3 narratives).
Metrics: 16 normalized behavioral features, including action ratios (read/talk/test/move), exploration efficiency, and decision consistency.

Key Contributions

Identification of Narrative Priors: The paper defines and isolates narrative priors—systematic behavioral tendencies activated by a task's story framing, independent of its decision structure. These priors are shown to be implicit biases inherited from pretraining corpora.
Quantification of Variance: Through variance decomposition (ANOVA) and classification, the study demonstrates that narrative framing explains 5–31× more behavioral variance than the assigned persona for information-gathering actions.
Behavioral Anchors: The authors identify behavioral anchors as the mechanism for cross-narrative transfer. These are persona descriptions containing concrete action words (e.g., "conversation" mapping to the talk action) that bind directly to the shared action space.
Causal Validation: Through intervention experiments, the paper causally confirms that removing anchor words from a persona reduces cross-narrative consistency by 95%, proving that abstract psychological descriptions are insufficient for portable behavior.
Generalization Framework: The structural isomorphism framework is validated on a held-out fourth narrative (Cooking Kitchen) and used to develop a persona-selection method that improves cross-narrative transferability.

Key Results

1. Narrative Dominance

Task narrative is the primary driver of agent behavior.

Variance: Narrative explains significantly more variance than persona (5–31×) for read, talk, and test actions.
Classification: Models can predict the narrative with 99.7–100% accuracy based on behavior, whereas persona classification accuracy is only 24.6–41.3% (barely above chance).
Domain Stereotypes: Agents adopt domain-stereotypical behaviors regardless of the persona:
- Medical (CI): High talk ratio.
- IT (SD): High read ratio.
- Crime (MM): High test ratio.
Performance Impact: These narrative-biased behaviors are often negatively correlated with task success. For example, the high talk bias in the medical domain correlates negatively with success in two of three models.

2. Persona Transfer and Consistency

Narrative-Confined Effects: While personas shape behavior significantly within a fixed narrative (explaining up to 64% of variance in some cases), their influence collapses across narratives. Cross-narrative transfer classification averages 10.4%, barely above the 10% chance baseline.
Behavioral Consistency Index (BCI): The authors introduce BCI to measure which personas maintain recognizable signatures across narratives. Only 5 of 10 personas maintain positive BCI across all models.
The Anchor Mechanism: High-BCI personas possess behavioral anchors. Regression analysis shows a strong correlation ( $R^2 = 0.65$ $R^{2} = 0.65$ ) between the number of anchors and BCI.
- Example: "Social collaborative" (anchors: "conversation," "discussion") maintains high consistency. "Big picture conceptual" (no anchors) has negative consistency.

3. Intervention Experiments

Causal Role of Anchors: Rewriting the "social collaborative" persona to remove concrete action words (replacing "conversation" with abstract "epistemic exchange") reduced BCI by 95% (from +0.22 to +0.01).
Predicting New Domains: The framework successfully predicted behavioral priors in a held-out "Cooking Kitchen" domain, confirming that narrative priors reflect pretraining associations rather than task-specific learning.
Robustness: Narrative priors persist even when:
- Action verbs are replaced with abstract codes (e.g., "ACTION ALPHA").
- Explicit behavioral directives (e.g., "perform 60% talk actions") are enforced. In some cases, the narrative prior (e.g., the forensic testing bias in MM) completely overrides the directive.

Significance and Claims

The paper argues that persona prompting is not a context-invariant control interface. The story wrapped around a task can shape agent behavior more strongly than the persona itself.

Reliability Challenge: Deploying the same persona across domains is risky because the narrative-induced behavior may actively degrade performance.
Design Principle: For behavioral control that survives task changes, personas must be grounded in concrete actions rather than abstract traits. Descriptions must map directly to the available action space (e.g., "prioritize reading documents" vs. "be thorough").
Methodological Contribution: The structural isomorphism framework provides a diagnostic tool to measure narrative influence, and the behavioral anchor analysis offers a practical path toward designing robust, portable agent behaviors.

The authors conclude that understanding and mitigating narrative priors is essential for building LLM agents with reliable behavior across diverse domains.