CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
CERA-MoA is a reinforcement learning framework where query routing and agent fine-tuning co-evolve, outperforming static Mixture-of-Agents routing baselines.
The paper (arXiv 2609.18779) introduces CERA-MoA, an iterative reinforcement learning framework in which a dynamic query router and independent agent policies co-evolve during post-training. A predictive familiarity estimator uses mid-layer hidden states to score agent competence without full rollouts, and cumulative-threshold adaptive routing activates a minimal tailored agent subset. Experiments across various domains show it outperforms state-of-the-art static-agent routing and fixed-workflow fine-tuning baselines.
- Predictive familiarity estimator uses mid-layer hidden states to score agent competence, avoiding full rollout overhead.
- Cumulative-threshold adaptive routing activates only a minimal tailored agent subset, balancing performance and efficiency.
- Targeted training-sample allocation based on evolving competence drives capability differentiation among agents.
Full article153 words · extracted from huggingface.co · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.18779