Navigating the Energy Landscape of Collaboration: Multi-Agent Communication Graph Generation via Score-Based Diffusion

The collective intelligence of Large Language Model (LLM)-based Multi-Agent Systems (MAS) is fundamentally governed by the underlying communication graph. However, discovering task-adaptive structures within this combinatorial search space remains a significant challenge. Existing methods, ranging from heuristic pruning to autoregressive generation, often lack a unified theoretical framework to guide the self-organization of agents into efficient teams. In this paper, we bridge non-equilibrium thermodynamics and generative modeling to formalize multi-agent graph generation as an energy minimization process. Specifically, we frame the emergence of efficient collaboration as a thermodynamic "cooling" process, where initially stochastic interactions converge to a low-energy, structured equilibrium. To implement this, We propose MAGE (Multi-Agent Communication Graph Generation), a score-based diffusion framework that constructs communication graphs by navigating the energy landscape via iterative denoising and first-order gradient guidance. Extensive experiments on representative benchmarks demonstrate that MAGE achieves state-of-the-art performance. Furthermore, qualitative analysis reveals that the generated graphs mirroring the functional specialization of human organizations, validating our thermodynamic hypothesis.

Imagine a team of AI agents—each specialized in different tasks like math solving or code generation—working together on a complex problem. They need to communicate, but if every agent talks to everyone else, the system becomes chaotic, expensive (high token cost), and inefficient. This paper proposes a method that views the agent team as a physical system where "energy" measures communication disorder: high‑energy states correspond to over‑connected, redundant graphs, while low‑energy states represent streamlined, efficient topologies. The proposed framework, MAGE, treats communication graph generation as an energy minimization process. Instead of blind search or random perturbation (zeroth‑order methods), MAGE exploits the duality between the generative score and the negative energy gradient. This allows it to follow first‑order gradients that point downhill on the energy landscape, guiding the reverse diffusion process toward low‑energy manifolds and eventually to the global optimum. Experiments on six benchmarks (math reasoning and code generation) show that MAGE achieves state‑of‑the‑art performance with the lowest token and financial cost. Visualizations confirm that MAGE successfully escapes high‑energy states and generates rational, task‑adaptive communication graphs. This approach offers a principled and efficient direction for designing collaborative multi‑agent systems.