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

TL;DR

This work investigates whether narrative priming can steer LLM agents toward collaboration in a finitely repeated public goods game. By contrasting homogeneous vs heterogeneous story prompts and examining scaling and robustness to selfish agents, the study shows that common narratives promote cooperation, while differing narratives can invert this effect in mixed populations. The results highlight a nuanced, context-dependent potential for narrative cues to shape multi-agent interactions, with implications for AI alignment and system design. While the findings are not a claim of human-like priming in LLMs, they reveal how narrative context can influence strategic behavior and coordination in AI agents, warranting careful consideration in safety and governance strategies.

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.

Figures (6)

• Figure 1: Repeated multi-round public goods game among homogeneous and heterogeneous LLM agents primed with various narratives.
• Figure 2: Violin plot of collaboration scores for homogeneous agent groups ($N=4$). Blue-shaded stories represent baseline conditions, while pink-shaded stories indicate meaningful narratives. The gray trend line represents the mean.
• Figure 3: Scaling experiment results for homogeneous agents across different group sizes. The ranking remains relatively consistent if we increase the group size.
• Figure 4: Collaboration scores in the robustness experiment ($N=4$) where one agent consistently contributes zero tokens. Overall cooperation compared to the baseline experiments decreases.
• Figure 6: Collaboration scores for homogeneous agent groups ($N=16$). Baseline conditions (blue) tend to yield lower collaboration, while meaningful narratives (pink) generally foster higher cooperation.