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

Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration

ICLR 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 Graph-of-Agents (GoA), a test-time multi-agent framework that samples relevant LLM agents via model cards, builds a relevance-weighted directed graph from peer scoring, runs bidirectional message passing, and pools outputs, achieving top results with 3 agents versus 6-agent baselines.

Problem

Orchestrating a large zoo of LLMs at test time suffers from selecting irrelevant agents, inefficient all-to-all communication, and costly response integration (token concatenation) as agent pools grow.

Approach

Model the multi-agent system as a query-specific directed graph: meta-LLM node sampling from model cards, peer-scored relevance edges with pruning, source-to-target then target-to-source message passing, and graph pooling (max or mean) to output a single answer; shown to generalize MoA.

Key results

• GoA-Max with 3 agents reaches 79.18 on MMLU and 54.78 on MMLU-Pro, exceeding 6-agent Self-MoA (78.14 and 54.19).
• GoA-Mean with 3 agents achieves 73.12 on MATH and 40.54 on GPQA, the best among multi-agent methods reported.
• On MMLU-Pro, GoA-Max uses 11 calls and 19.18k tokens versus MoA's 19 calls and 56.05k tokens, with higher accuracy (54.78 vs 53.33).

Surprising findings

• Reversing the message-passing direction causes the largest ablation drop (-2.60 on MMLU-Pro, -5.05 on GPQA).
• With gpt-4o, DyLAN (8 agents) achieves the best GPQA score (58.89), above GoA's 3-agent setting.

Limitations

• Relies on LLM-produced peer scores and meta-LLM selection, so quality depends on the judging and routing models.
• Evaluation uses stratified subsets for MMLU and MMLU-Pro and 100 samples for MedMCQA with gpt-4o, limiting statistical strength.
• Mostly 7-8B agent pool; generalization to larger or more diverse pools is only partially tested.

Experimental setup

Datasets: MMLU MMLU-Pro GPQA MATH HumanEval MedMCQA

Benchmarks: MMLU MMLU-Pro GPQA MATH HumanEval MedMCQA

Metrics: accuracy LLM calls tokens latency

Baselines: Debate Self-Consistency Refine ReConcile MoA Self-MoA DyLAN gpt-4o single model

Builds on: Mixture-of-Agents Chain-of-Thought DyLAN ReConcile Graph neural network message passing

Closest papers by meaning

• GPTSwarm: Language Agents as Optimizable Graphs (ICML 2024 · similarity 0.91)
• MOC: Multi-Order Communication in LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.90)
• HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems (ICML 2026 · similarity 0.90)
• Stochastic Self-Organization in Multi-Agent Systems (ICLR 2026 · similarity 0.89)
• AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.89)
• Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.89)