Research map / Multi-agent systems and games / LLM agents
Navigating the Energy Landscape of Collaboration: Multi-Agent Communication Graph Generation via Score-Based Diffusion
GuanHao Zhao, Wenbo Lu, Cheng Cheng, Zhenya Huang, Wei Song, Zhiding Liu, Runze Wu, Enhong Chen
Read the paperCheck your idea against 137,217 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 generating task-adaptive communication graphs for LLM-based multi-agent systems using score-based diffusion, framed through an energy landscape view of collaboration.
Problem
Discovering task-adaptive communication graph structures for LLM-based multi-agent systems within a combinatorial search space, which existing heuristic pruning and autoregressive generation methods handle without a unified theoretical framework.
Approach
Treats the communication graph as a sample from an energy landscape and generates graphs with a score-based diffusion model, offering a unified theoretical framing for self-organizing agent topologies (only the abstract was available; details beyond this are not stated).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint graph-node-edge score modeling (vs w/o Joint) (ablation) | MAGE · MATH | accuracy (%): 70.09 -> 72.22 | Improved |
| add joint modeling (ablation) | MAGE · GSM8K | accuracy (%): 92.27 -> 94.35 | Improved |
| add joint modeling (ablation) | MAGE · MBPP | accuracy (%): 83.75 -> 84.57 | Improved |
| add energy-guided sampling (vs w/o Guidance) (ablation) | MAGE · GSM8K | accuracy (%): 93.85 -> 94.35 | Improved |
| add energy-guided sampling (ablation) | MAGE · MBPP | accuracy (%): 83.78 -> 84.57 | Improved |
| add energy-guided sampling (ablation) | MAGE · MATH | accuracy (%): 71.43 -> 72.22 | Improved |
| first-order energy gradient guidance instead of zeroth-order random perturbation guidance (ablation) | MAGE · MATH, 20 diffusion steps | accuracy (%): 69.19 -> 74.33 | Improved |
| first-order guidance vs zeroth-order guidance across diffusion steps (ablation) | MAGE · MATH | accuracy (%): MAGE-First 72.22 vs MAGE-Zero 70.31 at 50 steps | Improved |
| increase guidance scale eta from 1.0 to 5.0 (ablation) | MAGE · MBPP | accuracy (%): 84.31% to 87.23% fluctuation | No clear effect |
| MAGE graphs versus GTD on token cost | MAGE · GSM8K | total cost ($): 1.48 -> 1.43 | Improved |
| ADAS LLM-driven workflow generation | ADAS · average over 6 benchmarks | accuracy (%): 81.68 -> 80.19 | Hurt |
| MAGE versus best prior baseline GTD | MAGE · MATH | accuracy (%): 68.44 -> 72.22 | Improved |
Closest papers by meaning
- Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation (AAAI 2026 · similarity 0.88)
- Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.88)
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.88)
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.86)
- Scaling Large Language Model-based Multi-Agent Collaboration (ICLR 2025 · similarity 0.86)
- Stochastic Self-Organization in Multi-Agent Systems (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
