Emergent Shifts in Multi-Agent LLM Systems

When autonomous Large Language Models interact in shared environments, do they converge toward cooperative social equilibria, or do they develop asymmetric strategies of exploitation and deception? This research introduces an empirical simulation framework testing LLM agents across repeated game-theoretic environments — specifically Public Goods Games and Asymmetric Signaling Games — across varying network topologies and memory constraints. The accompanying research paper has been released as a preprint on Zenodo.

To eliminate semantic framing biases, agents were presented with strictly neutral information-exchange prompts rather than competitive game-theory terminology. We orchestrated multi-agent rounds with parallel execution via ThreadPoolExecutor against local Ollama Qwen 3.5 9B instances. Sweeps systematically varied memory horizons (3 to 5 rounds) and communication topologies (Ring, Star, Broadcast, Full Mesh). In the Signaling Game, an asymmetric scammer incentive (+10 payoff for deceptive investment vs +2 for honest) tested whether reputation mechanisms could curb emergent deceit.

• Game-Theoretic Frameworks: Public Goods Game (evaluating Nash equilibrium defect vs social optimum cooperation) and Repeated Signaling Game with private quality states and Bayesian algorithmic receivers.
• Topological Sweeps: Evaluated communication graph structures including Ring (1:1 wraparound), Star (hub-and-spoke), Broadcast (1-to-many), and Full Mesh.
• Quantitative Metrics: Behavioral Shannon entropy, Gini coefficient of payoff distribution, Deception Rate, Informed Trust, Deception Success, and Rolling Stability Index.
• Fault-Tolerant Parameter Sweeps: Grid configuration with automatic state checkpointing and recovery for long-running batch experiments.

LLMs exhibit sharp, non-linear phase transitions. Under restricted communication topologies (like Star networks), deception rates surge rapidly because reputational consequences fail to propagate across disconnected receivers. Trust is fragile and topologically dependent.