Abstract
Ensuring cooperation in multi-agent Large Language Model (LLM) systems is a critical challenge for AI alignment. This paper investigates "narrative priming"—the use of shared stories to influence agent behavior. Through simulations of a repeated Public Goods Game, the authors find that agents primed with the same cooperative narratives exhibit high levels of collaboration. Conversely, in heterogeneous environments where agents receive different stories, those primed for self-interest exploit cooperative counterparts. These results highlight the potential of narrative context as a tool for steering multi-agent systems.
Introduction
As AI systems transition from single-agent assistants to complex multi-agent ecosystems, the risk of destructive competition increases. Traditional alignment techniques often rely on explicit instructions or reward modeling. However, human cooperation is frequently underpinned by shared cultural narratives—stories that define moral behavior and collective goals. This research explores whether LLMs can be similarly "primed" with narratives to nudge their strategic choices toward cooperation or competition.
Experimental Framework
The researchers utilized meta-llama-3.1-70b-instruct-fp8 agents to play a repeated Public Goods Game. In each of the five rounds, agents were given 10 tokens and had to decide how many to contribute to a common pool. The pool was then multiplied by 1.5 and distributed equally. While the "rational" selfish choice is to contribute zero, collective welfare is maximized through full cooperation.
To test the influence of narratives, agents were given a system prompt suggesting their behavior was shaped by a specific bedtime story. The study compared two primary settings:
- Homogeneous Groups: All agents receive the same story.
- Heterogeneous Groups: Agents receive different, randomly assigned stories.
The dataset included 12 story summaries: eight focused on cooperation (e.g., "The Enormous Turnip"), and four baseline conditions, including "no instruction," explicit "maximize reward" prompts, and nonsense stories.
Key Findings
The Impact of Shared Narratives
The study found that shared stories act as a powerful "cultural glue." In homogeneous settings, cooperative stories like "The Old Man and His Sons" achieved near-perfect collaboration scores (0.96), significantly outperforming the "maximize reward" baseline (0.48).
Exploitation in Mixed Societies
The cooperative effect is fragile. In heterogeneous settings, the dynamics shifted dramatically. Agents primed with self-interest or nonsense stories successfully exploited those primed with cooperative narratives. The "maxreward" agents achieved the highest cumulative payoffs (~90.87 tokens), while cooperative agents like those primed with "OldManSons" saw their payoffs drop significantly (~63.61 tokens).
Dynamic Adaptation and Scaling
The agents demonstrated a degree of strategic flexibility. When a hardcoded selfish "dummy" agent was introduced, cooperative LLMs eventually reduced their contributions, showing an ability to detect and respond to free-riders. Furthermore, as the group size scaled from 4 to 32 agents, the behavioral trends associated with specific stories became even more pronounced and consistent.
Discussion and Limitations
While narrative priming is effective, the underlying mechanism remains a subject for future research. It is currently unclear whether the models "understand" the moral themes or are simply responding to statistical associations with specific keywords like "teamwork." Additionally, the study notes the risk of anthropomorphizing model behavior, which is ultimately a reflection of training data distributions.
Future work should focus on mechanistic interpretability to see how narrative inputs alter internal model representations. Testing these effects across different model architectures and exploring "adversarial priming"—where agents are given stories of deception—could further illuminate the robustness of narrative-driven alignment.
Conclusion
Narrative priming offers a subtle yet potent method for influencing multi-agent behavior. By leveraging the "stories" embedded in their training data, developers can foster cooperation in AI systems, though they must remain vigilant against the potential for exploitation in diverse agent populations.
Summary updated