Why More Capable LLM Agents Can Make Financial Systems Riskier

Summary

Large language models are increasingly used in consequential systems, but a new study argues that improving individual model capability can sometimes worsen system-level outcomes. The researchers hypothesize that shared training and architectures make more capable models behave more similarly, so their actions become correlated rather than diversified. They develop a framework in which this correlation creates a non-diversifiable floor for system risk, then test it with an agent-based financial-market simulation using LLM traders with different levels of general-purpose capability. Frontier models show significantly more correlated behavior as capability increases. When the agents’ shared reasoning is accurate, adding more agents reduces market-level risk. When they operate in a common misinformation environment, however, the same correlation amplifies the downside and becomes a liability. The authors call this a capability paradox: better individual models do not necessarily produce better collective outcomes. Whether the pattern extends beyond financial markets remains an open empirical question.