Co-designing MoE Expert Selection for Agentic RL

Research shows that MoE expert routing naturally aligns with agentic operations like READ or UPDATE, suggesting that standard RL ignores this specialization and limits inference efficiency.