{
  "abstract": "Multi-agent language systems are often built as hand-designed workflows, where agents are assigned semantic roles and communication protocols are specified in advance. We propose NeuroMAS, a method that first treats a multi-agent language system as a trainable and scalable neural-network-like architecture with LLM agents as nodes and intermediate textual signals as edges. In NeuroMAS, agent nodes are role-free but structure-aware: the topology only determines how information can flow in general, while reinforcement learning training determines how nodes communicate, specialize, and coordinate. This formulation shifts multi-agent design from workflow engineering toward architecture design, where depth, width, connectivity, and growth protocol become scalable sources of capability. Further, we provide a theoretical perspective showing why such modular textual computation is more parameter-efficient when tasks admit hierarchical decompositions. Experiments show that NeuroMAS improves significantly over both inference-time and trained multi-agent baselines. We further find that organizational scaling is path-dependent: larger systems can be challenging to train from scratch, but become feasible when grown progressively from smaller trained systems. These results suggest that learned neural multi-agent systems are a promising scaling axis for LLMs.",
  "abstract_shown": true,
  "affiliations": [],
  "article_type": "Journal reference: The First Conference on Statistics and Trustworthy AI for Cross (X)-Domain Acceleration (2026)",
  "arxiv": "2605.16757",
  "author_total": 4,
  "authors": [
    {
      "name": "Haoran Lu"
    },
    {
      "name": "Luyang Fang"
    },
    {
      "name": "Wenxuan Zhong"
    },
    {
      "name": "Ping Ma"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/83ksx2e8nwwc1.bib",
  "cited_by": 0,
  "doi": "10.48550/arxiv.2605.16757",
  "fulltext_url": null,
  "id": "83ksx2e8nwwc1",
  "issns": [],
  "landing_url": "https://arxiv.org/abs/2605.16757",
  "language": null,
  "licence": "LicenseRef-arXiv-nonexclusive-1.0",
  "licence_label": "arXiv licence",
  "licence_source": "article",
  "licence_url": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
  "metadata_licence": "CC0-1.0",
  "native_subjects": [
    [
      "arxiv",
      "cs.AI"
    ],
    [
      "arxiv",
      "cs.MA"
    ],
    [
      "arxiv",
      "stat.ME"
    ],
    [
      "arxiv",
      "stat.ML"
    ],
    [
      "openalex",
      "1702|Multi-Agent Systems and Negotiation"
    ]
  ],
  "openalex": "W7161721520",
  "pdf_url": "https://arxiv.org/pdf/2605.16757",
  "pmcid": null,
  "pmid": null,
  "published": "2026-10-23",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/83ksx2e8nwwc1.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:35:25Z",
      "native_id": "2605.16757",
      "source": "arxiv",
      "url": "https://arxiv.org/abs/2605.16757"
    }
  ],
  "subjects": [
    {
      "field": "mathematics",
      "origin": "arxiv:stat.ME",
      "subfield": "statistics"
    }
  ],
  "title": "NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning",
  "title_text": "NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning",
  "type": "preprint",
  "updated_at": "2026-10-30T12:35:25Z",
  "url": "https://research.draconitic.com/paper/83ksx2e8nwwc1",
  "venue": "arXiv",
  "versions": []
}