China's HACPO Method Cuts AI Training Costs 50% in Multi-Agent Learning

⬤ Chinese research teams have introduced Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a training framework that lets diverse AI agents share experiences during training while staying fully independent at deployment. The approach tackles a persistent inefficiency in reinforcement learning: agents traditionally learn in isolation and never benefit from each other's breakthroughs. The system fits into a growing ecosystem of cooperative tools, including AgentScope open-source framework for building multi-agent AI systems.

⬤ At the heart of the framework is the HACPO algorithm, which enables structured sharing of rollout data across heterogeneous agents. Agents exchange successful trajectories to maximize sample efficiency and enable cross-agent knowledge transfer. Unlike older one-way distillation methods, HACPO introduces bidirectional learning so all agents improve at once while keeping their individual capabilities intact -- a meaningful leap over siloed optimization approaches that still dominate most production pipelines today.

⬤ The performance numbers are hard to ignore. HACPO outperforms baseline methods such as GSPO by an average of 3.3% while using roughly half the rollout cost, according to reported benchmarks. Reusing experiences across agents cuts redundant computation and wasted cycles -- a fix that lands at exactly the right moment, given that 79% of enterprises now deploy AI agents beyond pilot stages.

⬤ HACRL signals a broader shift toward cooperative AI architectures that prioritize shared learning over siloed optimization. Agents collaborate during training but act independently in production, which improves scalability across complex environments. This software momentum is running alongside hardware breakthroughs: Tesla is now completing AI chip design cycles in just 9 months - a sign that both layers of the AI stack are being rebuilt simultaneously.