CERA-MoA: A Mixture-of-Agents Framework Where Routing Mechanisms and Continually Learning Agents Co-Evolve
This work introduces CERA-MoA, an iterative reinforcement learning framework in which a dynamic router and multiple independent agent policies co-evolve: the router uses a predictive familiarity estimator built on mid-layer hidden states to assess each agent's semantic competence before generation, applies cumulative-threshold adaptive routing to activate a minimal agent subset, and proactively allocates targeted training samples based on agents' evolving competence, outperforming static-agent routing and fix-workflow fine-tuning baselines across diverse domains.