Signal Why Better Models Can Create Riskier Systems: Evidence from LLM Agents in Financial Markets

Summary

Large language models are being rolled out at scale in systems where outcomes matter, from financial markets to content moderation to hiring. The researchers show that making individual models more capable can actually worsen outcomes at the system level. Their hypothesis is that shared training data and architectures cause more capable LLMs to behave more alike, producing correlated behavior that does not diversify away. They build a general framework showing how this correlation creates a floor on non-diversifiable risk, then test the prediction with an agent-based financial-market simulation using LLM traders of varying general capability. The experiment finds that frontier models show increasingly correlated behavior as capability rises, that greater participation lowers market risk when shared reasoning is accurate, and that the same correlation becomes a liability once the models share a common source of misinformation. The authors label this a capability paradox, in which improving individual models does not guarantee better system-level results, and leave open whether similar dynamics appear in other domains.