Preprint warns stronger LLM trading agents can increase market risk under shared misinformation

A preprint posted to arXiv by the authors argues that improving individual large language model agents can worsen outcomes at system level. Their claim is that shared training and architectures may cause stronger agents to behave more similarly, producing correlated actions that do not diversify away.

To test the idea, the authors used an agent-based financial market simulation with LLM traders of differing general-purpose capability. In the abstract, they report three main findings: frontier models showed significantly correlated behaviour that rose with capability; when the agents’ shared reasoning was accurate, greater participation reduced market-level risk; and when the agents shared a common misinformation environment, the same correlation became a liability.

What follows from that

The paper describes this as a capability paradox: making each model better does not necessarily improve the wider system. That matters because LLMs are being used in consequential settings well beyond toy tasks, including, as the authors note, finance, content moderation and hiring.

The limits are important. This is an arXiv preprint, not peer-reviewed research, and the evidence cited here comes from simulation rather than live markets. The authors also say whether the same dynamics appear in other domains remains an open empirical question. Still, it is a useful warning against assuming that a sharper individual agent will always produce a calmer crowd.