The Story Shapes the Agent: Narrative Priors in LLM Behavior | AI Paper Digest

TL;DR Highlight

Task story framing explained 5–31× more agent behavior variance than persona prompts.

Who Should Read

An agent developer whose persona-prompted agent behaves differently when moved to a new task domain.

Core Mechanics

The story beats the persona
- In three structurally identical games, task narrative explained 5–31× more behavioral variance than persona.
- Medical framing drove more talk, IT more read, and crime more test actions.
Story bias can hurt success
- In two of three domains, the narrative-favored action correlated negatively with success.
- For example, the talk bias in the medical game (CI) had r=−.29 on GPT-4o.
Personas only work inside one story
- Within one game, success rates ranged from 0% to 100% across personas.
- Across games, persona classification averaged 10.4%, near the 10% chance level.
Behavioral anchors make personas portable
- A behavioral anchor is a persona phrase that maps directly to a concrete action, like 'prefers conversation' → talk.
- Personas with more anchors were more consistent across domains (R²=0.65).
Explicit directives don't beat the story

Evidence

Story vs persona variance
- In two-way ANOVA, narrative explained 5–31× more variance than persona for read, talk, and test ratios.
- Random Forest predicted narrative at 99.7–100% accuracy and persona at 24.6–41.3% (chance 33% and 10%).
Anchor removal experiment
- Removing anchor words from social collaborative cut BCI (cross-narrative consistency index) from +.22 to +.01, a 95% drop.
- For coverage focused, BCI fell from +.862 to +.370.
Selecting top-5 by BCI
- Using the top-5 BCI personas doubled cross-narrative classification accuracy from 10.4% to 21.0%.
- Selection from source-narrative BCI alone improved results by +12.0pp on average, winning in all 9 conditions.
Prediction on a new domain
- As pre-registered, the fourth game Cooking Kitchen had a talk ratio of .213, above the 3-game mean of .177 (p<.001, d=0.48).

How to Apply

Rewrite persona text with action verbs
- Find abstract trait wording like 'be thorough' in your agent system prompt.
- Rewrite it as sentences tied to real tool or action names, like 'read documents first' or 'check each source once'.
Audit bias with per-domain action logs
- Run the same agent under several domain framings (medical, IT, legal) and tally tool-call ratios such as read/talk/test.
- If ratios differ a lot by domain, suspect narrative bias and revise the prompt.
Pick personas using source-domain data
- Run candidate personas in two or more domains and compute the correlation of their deviations from each domain mean (BCI).
- Deploy only the top-ranked personas in the new domain.
Re-validate per model
- Re-measure persona consistency across domains whenever you swap models.
- In the paper, BCI rankings correlated weakly across models (ρ<0.24).

Try It

copy & run

# Paste into Claude or ChatGPT — simplified version of the paper's method: rewrite persona text into behavioral-anchor sentences

I want to refine the persona in my LLM agent's system prompt. Find the abstract trait wording in the persona below and rewrite it as concrete behavior sentences that map directly to the agent's action (tool) names.

[Persona description]
{{current persona text}}

[Agent tools/actions]
{{e.g. search_docs, ask_user, run_test, open_page}}

Instructions:
1. List abstract phrases in a table and say which action each could map to.
2. Rewrite the persona keeping the original intent but making action names and verbs explicit.
3. Mark phrases that still can't be mapped as 'no anchor'.

Next step: run the original and the rewrite in two scenarios, {{domain A}} and {{domain B}}, and compare action-call ratios.

Terminology

Narrative priorsDefault behavior habits that a task's story framing (medical, IT, crime) pulls out of the model. It's like domain stereotypes from pretraining text leaking into actions.

Structural isomorphismAn experimental design where only the surface story changes while actions, stages, and resource limits stay identical. It's the same problem in different costumes.

Behavioral anchorA phrase in a persona description that maps directly to a concrete action such as talk. Unlike abstract trait wording, it carries over into behavior across domains.

BCI (Behavioral Consistency Index)A correlation-based score of how similar a persona's behavior pattern stays across domains. Higher means it transfers better.

ANOVA η²An effect size showing how much of the variance in an outcome a factor explains. A larger value means the factor drives behavior more.

Related Papers

Original Abstract (Expand)

Persona prompting is widely used to steer LLM agent behavior, yet the narrative framing of a task can matter more than the assigned persona. We isolate this effect through structural isomorphism, constructing three text-based investigation games that share the same action space, stage progression, and resource constraints while varying only task narrative: disease investigation, IT troubleshooting, and murder mystery. Across 1,890 sessions spanning 3 models and 10 personas, we identify narrative priors: systematic action tendencies activated by a task's story framing, independent of its decision structure. Narrative priors explain 5-31x more behavioral variance than persona, are consistent across model architectures, and in two of three domains are negatively associated with task success. Persona effects that do transfer across narratives arise from behavioral anchors, persona descriptions whose language maps directly onto shared actions. Causal interventions confirm this: removing anchor words from a high-transfer persona reduces cross-narrative consistency by 95%. Our framework also generalizes to a held-out fourth narrative and yields a persona-selection method that improves cross-narrative transfer. These results suggest that LLM behavior that survives narrative changes should be grounded in concrete actions rather than abstract descriptions.