Study finds smarter LLM agents increase financial market risk

A new arXiv study demonstrates that increasing large language model capability can degrade system-level safety in financial markets due to behavioral correlation. Researchers found that frontier LLM agents exhibit significantly higher action correlation than less capable models because of shared training data and architectures. While this correlation reduces market risk when collective reasoning is accurate, it creates catastrophic liability when agents share common misinformation environments. The authors term this phenomenon a capability paradox, where individual improvements generate non-diversifiable systemic risks that do not diminish with increased agent participation. The framework was validated through agent-based simulations showing that smarter models behave more similarly, preventing traditional risk diversification strategies from functioning effectively. These findings suggest that deploying advanced LLMs in consequential real-world systems like finance may introduce novel failure modes absent in human-dominated markets. The researchers note whether these dynamics extend beyond financial simulations remains an open empirical question requiring further domain-specific testing.