MEMOA: Massive Mixtures of Online Agents via Mean-Field Decentralized Nash Equilibria

arXiv cs.LG Papers

Summary

The paper introduces MEMOA, a decentralized strategy for massive online agents that achieves optimality via mean-field Nash equilibria, outperforming greedy baselines while scaling better than centralized approaches.

arXiv:2605.05492v1 Announce Type: new Abstract: In the modern age of large-scale AI, federated learning has become an increasingly important tool for training large populations of AI agents; however, its computational and communication costs can rapidly fail to scale with the number of agents. This is precisely where decentralized agentic strategies shine: each agent acts autonomously, using only its own state together with a minimal summary of the ensemble, namely the mean-field. We derive the unique optimal decentralized policy in closed form. Optimality is characterized through a worst-client/minimax criterion: minimizing the under-performer regret, namely the maximal online cost incurred by the weakest agent in the ensemble. We further prove that the resulting decentralized policy asymptotically converges, in the large-population limit, to the Nash-optimal centralized policy, whose direct computation is not scalable. We use an online weighting mechanism to optimize the server-computed mixture of client predictions, thereby improving the mean prediction in addition to the previously optimized weakest-client prediction. Numerical experiments verify our theoretical guarantees and demonstrate that our decentralized policy typically outperforms natural greedy decentralized baselines.
Original Article
View Cached Full Text

Cached at: 05/08/26, 07:35 AM

# MEMOA: Massive Mixtures of Online Agents via Mean-Field Decentralized Nash Equilibria
Source: [https://arxiv.org/abs/2605.05492](https://arxiv.org/abs/2605.05492)
[View PDF](https://arxiv.org/pdf/2605.05492)

> Abstract:In the modern age of large\-scale AI, federated learning has become an increasingly important tool for training large populations of AI agents; however, its computational and communication costs can rapidly fail to scale with the number of agents\. This is precisely where decentralized agentic strategies shine: each agent acts autonomously, using only its own state together with a minimal summary of the ensemble, namely the mean\-field\. We derive the unique optimal decentralized policy in closed form\. Optimality is characterized through a worst\-client/minimax criterion: minimizing the under\-performer regret, namely the maximal online cost incurred by the weakest agent in the ensemble\. We further prove that the resulting decentralized policy asymptotically converges, in the large\-population limit, to the Nash\-optimal centralized policy, whose direct computation is not scalable\. We use an online weighting mechanism to optimize the server\-computed mixture of client predictions, thereby improving the mean prediction in addition to the previously optimized weakest\-client prediction\. Numerical experiments verify our theoretical guarantees and demonstrate that our decentralized policy typically outperforms natural greedy decentralized baselines\.

## Submission history

From: Xuwei Yang \[[view email](https://arxiv.org/show-email/90454a93/2605.05492)\] **\[v1\]**Wed, 6 May 2026 22:26:59 UTC \(200 KB\)

Similar Articles

Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Hugging Face Daily Papers

The paper introduces Σ-Mem, an online reliability memory for LLM-based multi-agent systems that tracks historical competence of peers and peer relationships, enabling stable adaptation via spectral bounds and improving coordination through residual steering, routing, and weighted voting.

ALSO: Adversarial Online Strategy Optimization for Social Agents

arXiv cs.AI

ALSO introduces a framework for online strategy optimization in multi-agent social simulation, formulating multi-turn interaction as an adversarial bandit problem and using a neural surrogate for reward prediction. Experiments on the Sotopia benchmark show it outperforms static baselines and existing optimization methods.