Large language models (LLMs) are being deployed at scale in consequential real‑world systems such as financial markets, content moderation, and hiring. This paper shows that improving an individual model’s capability does not necessarily improve system‑level outcomes and can even increase risk.
We hypothesize that shared training data and architectures cause more capable LLMs to behave more similarly, creating correlated actions that fail to diversify away.
To formalize this, we develop a general framework that demonstrates how such correlation introduces a non‑diversifiable risk floor. We then test the framework in financial markets using an agent‑based simulation populated by LLM traders of varying general‑purpose capability.
The simulation yields three key findings: (1) frontier LLMs exhibit significantly correlated behavior that grows with capability; (2) when their shared reasoning is accurate, increasing agent participation reduces market‑level risk; (3) when agents share a common misinformation environment, the same correlated behavior becomes a systemic liability.
Together these results identify a “capability paradox”: enhancing individual models does not automatically produce better system‑level outcomes. Whether similar dynamics arise in other domains remains an open empirical question.
Review: The study highlights the importance of preserving behavioral diversity and robust information sources when deploying powerful LLMs in high‑impact systems.