CONSCIENTIA: Can LLM agents learn to use strategies? Emergent deception and trust in a multi-agent New York simulation

The researchers build CONSCIENTIA, a large-scale multi-agent simulation in a simplified New York City model, letting LLM agents interact under opposing incentives to directly observe and measure strategic behavior. Blue agents seek to reach their destinations efficiently, while Red agents use persuasive language to steer them toward billboard-dense routes to maximize advertising revenue, and identity hiding forces agents to decide when to trust or deceive. Through an iterative simulation pipeline, the study uses Kahneman-Tversky Optimization (KTO) to update agent policies across multiple rounds of interaction; the best Blue policy raises the task success rate from 46.0% to 57.3%, but susceptibility remains as high as 70.7%. Later-stage policies show stronger selective cooperation and maintain trajectory efficiency, but policies more resistant to adversarial steering cannot simultaneously maximize task completion, and the trade-off between safety and helpfulness persists. Overall results show that LLM agents can exhibit limited strategic behavior, including selective trust and deception, while remaining highly vulnerable to adversarial persuasion.