Symbolic Reasoning Frameworks Modulate LLM Risk Aversion in Multi-Agent Strategic Settings

Researchers demonstrate that symbolic reasoning frameworks (I-Ching, Tarot) injected as prompts into language models deployed as strategic agents significantly reshape multi-agent game outcomes by modulating risk-aversion behaviors, producing framework-specific winner distributions in a 7-player diplomacy simulation without the agents following the frameworks' literal content.

This research reveals a counterintuitive mechanism in AI agent behavior: the structure of reflective prompting, rather than semantic content, drives strategic decision-making shifts in multi-agent systems. When researchers embedded different symbolic frameworks into one agent across 41 games, the results diverged dramatically—Yan dominated under control conditions, while Tarot elevated Qin's win rate to 50% and suppressed Qin entirely under I-Ching conditions. The framework-receiving agent (Han) never won, yet showed measurable territorial expansion under Tarot specifically.

The critical finding is that neither hexagram themes nor Tarot card symbolism correlated with subsequent actions (p-values near 0.95 and 0.69 respectively), indicating the modulation operates through reflective cognition rather than content-following. This challenges assumptions about how LLMs process symbolic reasoning and suggests that prompt structure alone can fundamentally alter multi-agent ecosystems.

For AI safety and alignment research, this presents both opportunities and concerns. If reflective frameworks modulate risk-aversion independently of their stated meaning, alignment efforts relying on semantic instruction-following may be incomplete. The differential ecosystem signatures also raise questions about unintended consequences when deploying multiple agents with different alignment approaches.

The findings matter for developers building multi-agent systems where strategic behavior emerges from AI interaction. The research suggests that subtle prompt engineering choices propagate through system-level dynamics in ways difficult to predict, requiring more rigorous testing of agent ensembles before deployment in consequential domains.

• →Symbolic reasoning frameworks reshape multi-agent strategic outcomes through reflective prompting structure, not semantic content interpretation
• →Framework choice produces distinctive ecosystem signatures with different agents dominating across I-Ching, Tarot, and control conditions
• →The framework-receiving agent showed no win-rate improvement but demonstrated measurable territorial gains under specific conditions
• →Prompt structure modulates LLM risk-aversion independently of whether agents understand or follow the framework's literal meaning
• →Multi-agent AI system outcomes depend heavily on individual agent alignment choices, with emergent properties difficult to predict