Mixed models still converged

The paper defines homogenization as agents converging to similar behaviors through their interactions, and treats it as a failure mode of multi-agent systems. It operationalizes the idea with three metrics: conformity to the majority, polarization toward extremes, and growing inertia against change across successive interactions. The authors are Prakhar Ganesh, Kyra Wilson, Luca Zappella, Barry-John Theobald, Nicholas Apostoloff, Lucas Monteiro Paes and Nivedha Sivakumar; they posted it on 7 October 2026.

They evaluate homogenization across code generation, hiring and scientific peer review. In code generation the paper says homogenization hides and amplifies correlated errors, which can create systemic vulnerabilities. In hiring it lets the influence of biased agents persist long after those agents are removed. In peer review it produces uneven evaluation standards across research areas.

The paper's conclusion is that multi-agent evaluations must look past aggregate performance to the dynamics of interaction. It tests two simple ways to add diversity, sampling stochasticity and mixed-model systems, and reports that neither reduces the homogenization risk. What would work is left open.