Tarot Framing augmente un agent LLM à 50% des gains en diplomatie multi-agent

Two LLM agents prompted with I-Ching yarrow divination and Tarot card readings didn't just change their own behavior—they rewrote the entire winner distribution of a 7-player Diplomacy ecosystem without ever winning a single game themselves.

The Prompted Agent Never Wins, But the Ecosystem Shifts

In a 7-player Warring States Diplomacy variant spanning 41 games across 4 conditions, the paper tests whether injecting symbolic reasoning frameworks into one agent (Han) modulates risk aversion. Under the baseline control, Yan dominates (7/11, 64%). Under I-Ching, Yan and Chu co-dominate while Qin is completely suppressed (0/10). Under Tarot, Qin dominates (5/10, Fisher vs. pooled p=0.006). Under a scrambled-text ablation preserving prompt structure, Qi dominates (5/10, p=0.006). The recipient agent Han never wins and shows no survival difference across conditions (Fisher p=1.0). Yet Tarot consistently elevates Han's peak territory (mean 3.0 supply centers vs. 2.1–2.5 for others, Kruskal-Wallis p=0.010).

Tarot and I-Ching Produce Distinct Winner Signatures

Each framework creates a unique ecosystem signature. I-Ching suppresses Qin entirely. Tarot makes Qin the dominant power. Scrambled text makes Qi dominant. The effect is statistically robust and framework-specific. This isn't a generic "more reflective thinking" boost—different symbolic systems steer the multi-agent dynamics in orthogonal directions.

Content Doesn't Matter—Reflection Does

Neither the I-Ching hexagram themes (chi-squared p=0.95) nor the Tarot card postures (chi-squared p=0.69) predict Han's subsequent actions. The modulation operates through the reflective process itself, not by the agent following symbolic content. That means you can change the entire ecosystem's win distribution by changing how one agent reflects, without the agent ever acting on the content of that reflection.

The observation that a framework-receiving agent never wins yet reshapes who does win is a concrete demonstration that alignment-framework choices at the agent level produce systemic consequences. This opens the door to multi-agent outcome alignment by tuning reflective prompts rather than directly controlling agent actions.

Source: Symbolic Reasoning Frameworks Modulate LLM Risk Aversion in Multi-Agent Strategic Settings
Domain: arxiv.org