Co-evolving routing and learning improves efficiency in multi-agent systems
CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
Artificial IntelligenceMachine Learning
Summary
When multiple AI agents work together, deciding which agent should handle each task is important but usually separate from how the agents learn. This paper shows that by letting the routing system and the agents learn and improve together over time, the overall system can get better and faster. The authors created a method that predicts which agent is most suitable without doing long tests and routes tasks to only the necessary agents. This approach encourages agents to specialize in different tasks and outperforms older methods that treat routing and learning separately.
What this means in practice
- •For ai system developers: Build AI systems that dynamically assign tasks to specialized agents for better performance and efficiency during ongoing training.
- •For cloud service teams: Optimize resource use by activating only a tailored subset of AI agents based on real-time capability estimates.
Authors
Jiaxuan Jiang, Liyuan He, Zhixuan Fang
Abstract
Current Mixture-of-Agents (MoA) paradigms generally treat query routing and agent fine-tuning as separate processes, limiting their ability to respond to evolving agent capabilities. This disconnect prevents routing strategies from adapting to evolving agent capabilities during post-training and prevents agents from achieving synergistic data-driven specialization. To resolve this, we introduce CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents), an iterative reinforcement learning framework where the dynamic router and independent agent policies co-evolve. We design a predictive familiarity estimator that leverages mid-layer hidden states to evaluate semantic competence among agents, avoiding the overhead of full rollouts. Based on these familiarity scores, a cumulative-threshold adaptive routing mechanism dynamically activates a tailored minimal agent subset, achieving a trade-off between task performance and efficiency. By proactively allocating targeted training samples to agents based on their evolving competence, CERA-MoA promotes capability differentiation. Extensive experiments across various domains demonstrate that CERA-MoA outperforms state-of-the-art static-agent routing and fix-workflow fine-tuning baselines.