The Power of Stories: Narrative Priming Shapes How LLM Agents Collaborate and Compete

Abstract: According to Yuval Noah Harari, large-scale human cooperation is driven by shared narratives that encode common beliefs and values. This study explores whether such narratives can similarly nudge LLM agents toward collaboration. We use a finitely repeated public goods game in which LLM agents choose either cooperative or egoistic spending strategies. We prime agents with stories highlighting teamwork to different degrees and test how this influences negotiation outcomes. Our experiments explore four questions:(1) How do narratives influence negotiation behavior? (2) What differs when agents share the same story versus different ones? (3) What happens when the agent numbers grow? (4) Are agents resilient against self-serving negotiators? We find that story-based priming significantly affects negotiation strategies and success rates. Common stories improve collaboration, benefiting each agent. By contrast, priming agents with different stories reverses this effect, and those agents primed toward self-interest prevail. We hypothesize that these results carry implications for multi-agent system design and AI alignment.

• The paper demonstrates that cooperation-themed narrative priming significantly increases collective contributions in public goods games, achieving near-perfect collaboration in homogeneous settings.
• It employs iterated negotiation games with bootstrapped statistical analysis across varying agent counts to validate the impact of narrative prompts.
• The study highlights that in heterogeneous groups, contrasting narratives can reverse cooperative benefits, offering critical insights for multi-agent system design and AI alignment.

Narrative Priming as a Control Variable in LLM Multi-Agent Cooperation

Introduction and Motivation

The paper "The Power of Stories: Narrative Priming Shapes How LLM Agents Collaborate and Compete" (2505.03961) explores narrative priming as a method to modulate collaboration and competition in populations of LLM agents. Motivated by the social science hypothesis that shared narratives underpin large-scale human cooperation, the study evaluates whether analogous mechanisms can be leveraged to align LLM agents in multi-agent negotiation games. The central investigation targets both the practical effects of narrative priming and its theoretical implications for AI alignment and agent strategy adaptation.

Experimental Framework and Methodology

The research employs a repeatedly iterated public goods game as its testbed, adopting classical game-theoretic mechanics: each of $N$ agents receives $T=10$ tokens per round and must decide how many to contribute to a shared pool; the pooled sum is multiplied ($m=1.5$) and redistributed equally. Agents only have post-hoc access to other agents’ contributions, and the dominant single-shot strategy is non-cooperation ($t=0$). However, in the multi-round setting, agents can reciprocate and adaptively adjust their strategies, thus simulating negotiation and collective adaptation.

Narrative priming is operationalized by embedding each agent’s initial prompt with a story—either cooperation-themed (eight curated narratives) or non-cooperative / nonsensical baseline (four controls). In the homogeneous setting, all agents receive the same narrative; in heterogeneous trials, stories are randomly assigned. The experimental design encompasses:

• Homogeneous groups (shared story priming),
• Scaling with variable agent counts ($N \in \{4, 16, 32\}$),
• Robustness trials with a persistently selfish agent (contributes zero tokens),
• Heterogeneous groups (mixed narrative priming).

Collaboration scores, defined as the ratio of actual cumulative contributions to maximum possible contributions, serve as the key quantitative outcome metric. Statistical robustness is ensured via repeated trials and bootstrapped confidence interval analyses.

Key Results and Quantitative Analysis

The empirical results establish several robust claims:

1. Strong Narrative Effect in Homogeneous Groups: Narrative priming with cooperation-centered stories consistently yields higher collaboration scores. For example, stories such as "OldManSons" or "Turnip" result in mean scores close to 1.0 ($0.96 \pm 0.05$, $0.95 \pm 0.04$ for $N=4$), indicating near-perfect agent cooperation. Baseline or egoistic narratives yield substantially lower cooperation ($0.48 \pm 0.12$ for "maxreward").
2. Scaling Consistency: Increasing $N$ to 16 or 32 preserves the ordering and magnitude of narrative effects, with meaningful stories sustaining high collaboration and baselines remaining low (e.g., "Turnip": $0.99 \pm 0.01$ for $N=32$).
3. Robustness to Free-Riding: Introducing an agent that always defects drops overall group collaboration but preserves the internal ranking of narrative prompts. Agents adaptively reduce contributions, but narratives with stronger cooperation themes still mitigate free-rider impact.
4. Contradictory Effect in Heterogeneous Populations: When agents are primed with different stories, self-interested narratives ("maxreward") dominate, yielding higher individual payoffs and suppressing overall cooperation. Cooperation-themed stories underperform in this competitive environment, reversing the effect observed in homogeneous settings.

Pairwise statistical comparisons with bootstrapped confidence intervals reinforce the reliability of these ranking effects, with very few comparisons failing to reach statistical significance.

Implications and Theoretical Impact

The paper advances several important implications:

Multi-Agent System Design and AI Alignment

The results highlight narrative priming as a critical control variable for emergent group behavior in LLM multi-agent systems. The magnitude and reversibility of the effect suggest that prompt engineering, narrative design, and contextual embedding can steer collective agent dynamics—offering a tractable route for automated alignment in decentralized settings.

Limits of Implicit Priming

A bold finding is that narrative effects only materialize when agents share a common story; mixed narrative populations default to competitive equilibria. This points to a nuanced understanding of alignment: mere broadcasting of collaborative themes is insufficient in heterogeneous systems, as competitive narratives can undermine global cooperation.

Adaptation and Strategy Formation

LLM agents demonstrate dynamic adaptation, not simply rigid heuristic responses. The ability to respond to adversarial (selfish) agents and alter strategies based on context elevates the generality of narrative priming beyond simple instruction following, suggesting a statistical learning process reminiscent of human priming effects.

Ethical and Environmental Considerations

The study addresses ethical vectors (reliability and moral context via priming) and environmental concerns, reporting detailed energy expenditure for execution ($44.8$ kWh over $\sim$32 hours, $334,000$ inference calls using LLaMa 3.1 70B parameter model).

Limitations and Recommendations for Future Research

Several open questions remain:

• Mechanistic interpretability of narrative priming effects: mapping narrative inputs to transformer attention and value pathways.
• Decay and persistence of priming across repeated games and time windows.
• Adversarial narrative priming: destabilization via malicious or deceptive prompts.
• Scaling laws and cross-model generalization: performance under different architectures, non-RLHF variants, and variable emotional valence/narrative structure.
• Alignment with human strategies—comparative studies to quantify overlap between LLM and human cooperative/competitive behaviors.

Methodological expansion could include systematic narrative selection, emotional and semantic annotation of story content, and direct measurement of prompt-induced representation shifts within LLMs.

Conclusion

This work establishes narrative priming as a decisive control mechanism for modulating LLM agent cooperation and competition in multi-agent systems. Shared narrative prompts reliably steer agents toward collaborative equilibria, while heterogeneous narrative assignment reverses the effect, favoring competitive strategies. The findings underscore the importance of narrative context for AI alignment, highlight strategic adaptation in agent behavior, and suggest both practical and theoretical avenues for leveraging narrative priming in designing robust and aligned multi-agent workflows. Further research should seek mechanistic explanations, test priming durability, and extend cross-model evaluations to deepen our understanding of narrative-driven agent dynamics in artificial collectives.