Frontier LLMs Exhibit Capability-Linked Correlation in LLM Trader Simulations

The arXiv paper 2609.04373 reports agent-based market simulations in which GPT-4-class and later models generate buy-sell decisions whose correlation rises from 0.31 at 7B scale to 0.68 at frontier scale. When agents receive identical inaccurate signals, this correlation converts accurate individual reasoning into synchronized position-taking that widens bid-ask spreads and amplifies drawdowns by 2.4 times relative to low-correlation baselines.

Data from 1,200 simulation runs show that adding more frontier agents reduces volatility only when environmental signals are veridical; the same increase raises tail-risk measures once misinformation is injected. The framework isolates a capability paradox: higher benchmark scores on individual tasks tighten behavioral alignment across models trained on overlapping corpora, eroding the diversification assumed in classical portfolio theory.

Operational deployment therefore requires explicit measurement of cross-agent decision overlap before scaling participation. Regulators monitoring AI-driven order flow lack current mandates to collect model-identity metadata, leaving the correlation channel invisible to existing circuit-breaker logic.

Next steps include live-market pilots that instrument trader-model provenance and test diversity constraints at the exchange level. Absent such controls, the risk floor identified in simulation transfers directly to venues where LLM agents already route institutional flow.