Our JMLR paper Heterogeneous-Agent Reinforcement Learning is selected as a Capital Frontier Academic Achievement
Selected by the Beijing Association for Science and Technology (Oct 2026). HARL moves cooperative multi-agent RL beyond parameter sharing to genuinely heterogeneous agents: built on the multi-agent advantage decomposition lemma and a sequential update scheme, it derives HATRPO / HAPPO and the Heterogeneous-Agent Mirror Learning (HAML) framework, whose algorithms provably improve the joint return monotonically and converge to Nash equilibrium.