Mesh Inference Lets AI Agents Reason Together Without a Central Brain — Here's How

Hongwei Xu's arXiv paper introduces Mesh Inference, a formal mathematical model for collective intelligence without any central coordinator. In this framework, independent agents each hold private state and communicate only through 'admitted, typed observations'—akin to asking questions and receiving answers—without ever sharing weights, gradients, or hidden states. The agents can span different teams, networks, or organizations, making the model relevant for federated and privacy-preserving AI. Xu models the mesh as a coupled free energy that each agent relaxes locally, inspired by the energy-minimizing nature of inference in neural networks.

Xu proves three major properties of this model. First, mesh inference converges to a unique answer for any admission policy, symmetric or not, because the coupling matrix is always an M-matrix. Second, it is 'identification-complete': it derives exactly the centralized optimum when the contributing views are carrier-connected (i.e., sufficiently informative). Third, it is 'observation-only'—no node transmits its internal state, and confidentiality is mathematically dual to identification. In the linear-Gaussian regime, every derived answer equals the centralized optimum at O(diam²) latency, which Xu calls 'the measured price of removing the center.' The paper leaves as an open problem whether the non-linear closure produces an upgraded answer or a confident error—a critical question for real-world deployment. This is, by Xu's account, the first formal model of mesh inference.