Research map / Multi-agent systems and games / LLM agents

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

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HieraMAS, a multi-agent system with supernodes (mixtures of heterogeneous LLMs per role) whose LLM selection is learned via multi-level rewards and whose topology is chosen by a GCN graph classifier over candidate DAGs.

Problem

Existing multi-agent system optimizers treat each agent as a monolithic unit and optimize topology, roles, or LLM routing in isolation, and joint optimization suffers from credit assignment problems at node and edge level.

Approach

Supernodes with propose-synthesis LLM mixtures per role; a two-stage method that first learns LLM selection with multi-level (node + final) rewards under random graphs, then trains a GCN classifier to pick a topology from a pre-generated DAG pool instead of per-edge optimization.

Key results

• HieraMAS achieves 94.61% average accuracy across HumanEval++, MATH and MMLU-Redux, the best average in the main table (Full-Graph 92.78%, AFlow 92.69%).
• On MMLU-Redux with GPT-5-Mini, HieraMAS costs $1.29 versus $4.23 for Full-Graph, and its training cost on HumanEval++ is reported as 18.41x cheaper than AFlow.
• Ablation on MATH: removing the graph scoring mechanism drops accuracy by 3.45 points (96.67 to 93.33).

Surprising findings

• The skip token is rarely selected, so the method reduces cost through sparser topologies rather than removing roles.
• Ablating LLM selection (all agents use GPT-5-Mini) yields slightly higher MATH accuracy (97.78 vs 96.67), though at higher cost.

Limitations

• Generalization to unseen MMLU subjects is weaker than some baselines (e.g., MASRouter drops to 56% with GPT-5-Mini; HieraMAS 68% vs AFlow/GDesigner 72%).
• Theoretical assumptions (e.g., q>0.5, bounded rewards, optimal graph in candidate pool) are heuristic and simplified.
• Results rely on a fixed pool of LLMs accessed via APIs and a small set of 3 benchmarks.

Experimental setup

Datasets: HumanEval++ MATH MMLU-Redux

Benchmarks: HumanEval++ MATH MMLU-Redux

Metrics: Pass@1 accuracy cost (USD)

Baselines: Base CoT Self-Consistency Self-Consistency+CoT LLM-Debate Full-Graph Random-Graph AFlow GDesigner MASRouter

Builds on: MASRouter GDesigner AFlow Mixture-of-Agents LLM-Debate

Closest papers by meaning

• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.90)
• AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.89)
• Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems (NeurIPS 2025 · similarity 0.88)
• Stochastic Self-Organization in Multi-Agent Systems (ICLR 2026 · similarity 0.88)
• MOC: Multi-Order Communication in LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.88)
• SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.88)