Researchers introduce CERA-MoA, a reinforcement learning framework for Mixture-of-Agents systems in which a dynamic query router and independently fine-tuned agent policies co-evolve, rather than being trained as separate static stages. A predictive familiarity estimator uses agents’ mid-layer hidden states to gauge semantic competence without running full rollouts, feeding a cumulative-threshold adaptive routing mechanism that activates a minimal, tailored agent subset per query and directs training data toward each agent’s evolving strengths. Across multiple domains, CERA-MoA outperforms state-of-the-art static-agent routing and fixed-workflow fine-tuning baselines on both task performance and efficiency.
