CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
CERA-MoA ties the router and the agents into one reinforcement-learning loop so both adapt as the agents keep learning.
The paper says existing Mixture-of-Agents setups usually route queries and fine-tune agents separately, which leaves routing stale as agent abilities change. Its proposed framework uses a familiarity estimator from mid-layer hidden states to judge which agents are semantically competent without running full rollouts. It then activates a minimal agent subset through cumulative-threshold routing, aiming to balance performance with efficiency. The authors report gains over static-agent routing and fixed-workflow fine-tuning baselines across multiple domains. HF Daily Papers' note
The paper says existing Mixture-of-Agents setups usually route queries and fine-tune agents separately, which leaves routing stale as agent abilities change. Its proposed framework uses a familiarity estimator from mid-layer hidden states to judge which agents are semantically competent without running full rollouts. It then activates a minimal agent subset through cumulative-threshold routing, aiming to balance performance with efficiency. The authors report gains over static-agent routing and fixed-workflow fine-tuning baselines across multiple domains. HF Daily Papers' note
score 4