HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems

A team of researchers (Tianjun Yao, Zhaoyi Li, Zhiqiang Shen) has proposed HieraMAS, a novel hierarchical framework that fundamentally rethinks how to build multi-agent systems (MAS) powered by large language models. The key innovation is moving beyond treating each AI agent as a single, indivisible LLM. Instead, HieraMAS introduces 'supernodes,' where each functional role (e.g., a coder, a planner) is implemented by a mixture of multiple, potentially heterogeneous LLMs (like GPT-4, Claude 3, or Llama 3) working together in a propose-synthesis structure. This allows the system to leverage specialized strengths from different models within a single agent. Concurrently, the framework optimizes the inter-node communication topology—how these superagents talk to each other—addressing two critical dimensions most prior work tackles separately.

Optimizing such a complex system presented a unique credit-assignment challenge, as final task success depends heavily on the underlying LLMs' capabilities, which can mislead reinforcement learning methods. HieraMAS solves this with a sophisticated two-stage algorithm: first, multi-level reward attribution provides fine-grained feedback at both the node and system level; second, graph classification treats selecting the optimal communication structure as a holistic decision, not just optimizing individual connections. Experiments on reasoning and coding benchmarks show HieraMAS substantially outperforms existing MAS approaches. The framework delivers not just higher performance but also a superior cost-performance trade-off, enabling more efficient use of expensive, high-capability models alongside cheaper, specialized ones.