CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kCERA-MoA (Qwen3-4B)#39Accuracy: 93.5
mathematical-reasoning-on-olympiadbenchCERA-MoA (Heterogeneous pool)#39Accuracy: 44.1
code-generation-on-humanevalCERA-MoA (Qwen3-4B)#51Pass@1: 82.5
code-generation-on-mbppCERA-MoA (Qwen3-4B)#88Accuracy: 60.3