CERA-MoA: Co-Evolving Routing Mechanisms with Conti... | AI Research

CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
Current multi-agent systems often struggle because they treat the "router" (the mechanism that decides which agent handles a task) and the "agents" (the models performing the task) as separate, static components. This disconnect prevents the system from adapting as agents learn new skills or as task requirements change. CERA-MoA addresses this by introducing an iterative reinforcement learning framework where the router and the agents co-evolve. By dynamically matching specific training queries to the agents best suited to solve them, the system encourages agents to specialize in different domains while the router simultaneously learns to recognize these evolving capabilities. The same large language models question is explored in Trace2Tower, which adds a research perspective.

How the System Learns and Specializes

The core of CERA-MoA is a closed-loop feedback process. When a query arrives, a "predictive familiarity estimator" evaluates which agents are most competent to handle it. This estimator uses intermediate hidden states from the language model to judge compatibility, which is much faster than running a full generation process to see if an agent succeeds. As agents receive and solve tasks, they improve their performance through reinforcement learning. Simultaneously, the router updates its understanding of each agent’s expertise. This creates a virtuous cycle: agents that perform well on certain types of queries are rewarded with more similar tasks, naturally pushing them to become domain specialists rather than generalists.

Efficient Task Routing

Instead of using a rigid system that always picks a fixed number of agents, CERA-MoA uses a "cumulative-threshold adaptive routing" mechanism. The router calculates a familiarity score for each agent and activates only the smallest subset of agents necessary to meet a predefined competence threshold. This approach allows the system to balance high-quality results with computational efficiency. During training, the system also includes exploration bonuses to ensure that agents are not "starved" of tasks, allowing them to continue learning and potentially developing new competencies over time. The ai agents story also surfaces in Google AI Introduces EnvHarness for Adaptive..., adding another angle.

Performance and Flexibility

Experiments show that CERA-MoA consistently outperforms existing methods that rely on static routing or fixed-workflow fine-tuning. By integrating the routing strategy directly into the learning loop, the framework achieves better results across a variety of domains. The system is also highly flexible; it can be implemented by sharing a single base model with independent adapters for each agent, or by using a pool of entirely different, heterogeneous base models. This adaptability allows the framework to scale effectively while maintaining the specialized reasoning capabilities of the individual agents. The ai agents story also surfaces in Stanford Researchers Develop TRACE to Fix..., adding another angle. as detailed in the full paper on Arxiv