The Story Shapes the Agent: Narrative Priors in LLM Behavior¶

🕒 Published (v1): 2026-07-20 23:03 UTC · Source: Arxiv · Venue: COLM 2026 · link

Why this paper was selected

COLM 2026; narrative framing empirically dominates persona in steering agent behavior

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

Task narrative framing—not persona prompting—is the primary driver of LLM agent behavior, explaining 5–31× more behavioral variance than the assigned persona across structurally identical tasks. This effect, termed "narrative priors," reflects domain stereotypes from pretraining and is negatively correlated with task success in most settings. Persona effects that do generalize across narratives are causally traceable to concrete action words ("behavioral anchors") in the persona description.

Problem¶

Practitioners rely on persona prompting to steer LLM agent behavior across task deployments, but it is untested whether persona-induced behavior transfers when the same agent operates in a different task domain. Prior work confounds decision structure with narrative surface, making it impossible to isolate the narrative's independent contribution to behavior.

Method¶

Three structurally isomorphic text-based investigation games are constructed—Crystal Island (disease), Sequential Diagnosis (IT), and Mansion Mystery (crime)—sharing the same 4 action types (read, talk, test, move), 4 stages, 7 scenarios, and resource constraints, differing only in surface narrative. Ten learning-style personas are run across all three games on three models (LLaMA-3.1-70B, GPT-4o, GPT-4o-OSS-120B) for 1,890 total sessions.

Behavioral effects are quantified via two-way ANOVA (\(\eta^2\), Narrative \(\times\) Persona) on 16 normalized behavioral features and Random Forest classifiers. Cross-narrative persona transfer is measured by a Behavioral Consistency Index (BCI)—the mean pairwise Pearson correlation of z-normalized behavioral feature vectors \(\hat{\mathbf{b}}_p^{(g)}\) across all narrative pairs:

\[\text{BCI}_p = \frac{1}{3}\sum_{(g_i,g_j)} \text{corr}\!\left(\hat{\mathbf{b}}_p^{(g_i)},\, \hat{\mathbf{b}}_p^{(g_j)}\right)\]

Causal confirmation uses a controlled rewrite: anchor words are removed from the "social collaborative" persona while preserving its psychological content, and the BCI change is tested via bootstrap, permutation, and Fisher-z.

Key Contributions¶

Introduces structural isomorphism as an experimental design for isolating narrative influence from task structure in LLM agent studies.
Documents narrative priors: pretraining-driven action biases (medical→talk, IT→read, crime→test) that are consistent across architectures and negatively correlated with task success in 2 of 3 domains.
Defines the Behavioral Consistency Index (BCI) to quantify persona transferability across narratives.
Identifies behavioral anchors (concrete action words in persona descriptions) as the causal mechanism for cross-narrative transfer; removing anchors collapses BCI by 95%.
Provides a persona-selection algorithm using source-narrative BCI that improves cross-narrative identifiability by +12.0 pp on average across all 9 model–target combinations.
Generalizes to a held-out fourth narrative (Cooking Kitchen), confirming prediction of its talk-ratio bias (\(t=5.64\), \(p<.001\)).

Results¶

Narrative explains 5–31× more variance than persona for information-gathering actions (read, talk, test) via ANOVA \(\eta^2\); all narrative effects \(p < 10^{-15}\).
Random Forest narrative classification: 99.7–100% accuracy; persona classification: 24.6–41.3% (chance 10% and 33% respectively).
Cross-narrative transfer (all 10 personas): 10.4% accuracy, barely above the 10% chance baseline.
Restricting to top-5 BCI personas doubles transfer accuracy to 21.0% (+11.6 pp under leave-one-narrative-out).
Anchor-count regression: \(R^2 = 0.65\), \(p < 10^{-7}\) (improves to \(R^2 = 0.68\) with 4 narratives).
Causal anchor intervention (social collaborative): BCI drops from +.22 to +.01 (95% reduction); \(\Delta\text{BCI}\) 95% CI \([-0.38, -0.04]\), permutation \(p = .034\).
Second persona (coverage focused) replication: BCI drops from +.862 to +.370 (\(\Delta = -0.492\), permutation \(p < 0.0001\), effect size 7.1 SD).
BCI predicts cross-narrative transfer (\(\rho = 0.68\)–\(0.76\), \(p<.05\)); simple action-word count baseline does not (\(\rho = 0.38\)–\(0.55\), \(p>.10\)).

Limitations¶

Only 3 (+ 1 held-out) narrative domains and 10 personas are tested; results may not generalize to less stereotyped or more novel narratives.
The action space is discrete and small (4 types); richer open-ended agent environments may exhibit different dynamics.
Models tested are limited to three; the finding that BCI is partially model-dependent (cross-model \(\rho < 0.24\)) means persona rankings must be re-derived per model family.
The mechanism for narrative priors is attributed to pretraining corpus statistics but is not directly verified (no corpus analysis).
Structural isomorphism is hand-engineered; scaling to real-world agentic tasks with differing decision structures remains an open challenge.

Relevance to Agentic AI / LLM Agents¶

This work directly challenges a core assumption in LLM agent engineering—that persona prompting provides portable behavioral control—by demonstrating that narrative framing dominates agent behavior through implicit pretraining biases. For anyone deploying LLM agents across task domains (e.g., customer service, coding, medical triage), the findings imply that surface-level task descriptions may undermine intended behavioral specifications in ways that are invisible within a single deployment context. The behavioral anchor mechanism provides an actionable design principle: agent steering prompts should use concrete action-grounded language rather than abstract trait descriptions to achieve robust cross-context transfer. The structural isomorphism framework also offers a reusable evaluation methodology for disentangling agent behavior from task structure in multi-agent and agentic benchmarks.