GENERAL EDGE‑CLOUD HETEROGENEOUS MULTI‑AGENT COLLABORATIVE REASONING ARCHITECTURE WITH LOCAL‑PRIVACY CONSENSUS MECHANISM
Keywords:
Edge‑cloud collaboration, Multi‑agent, Local privacy, Retrieval‑augmented generation, Heterogeneous reasoningAbstract
Multi‑agent systems for large‑language‑model tasks are evolving from single‑model generation toward complex‑task collaborative reasoning. Conventional schemes concentrate raw data, private knowledge and reasoning workflows on the cloud, and merge outputs via majority voting among homogeneous agents, which hardly satisfy native‑privacy protection, heterogeneous task division, private‑knowledge enhancement and traceable‑evidence requirements simultaneously. This paper proposes GEHMAR, a general edge‑cloud heterogeneous multi‑agent collaborative reasoning architecture oriented to local‑privacy consensus. The architecture decouples data, knowledge and computation: the edge undertakes private‑data parsing, local RAG retrieval, evidence compression and privacy isolation, while the cloud only receives desensitized evidence after minimal processing. A heterogeneous reasoning group consisting of Analysis Agent, Multimodal Verification Agent, Evidence Review Agent and Arbitration Consensus Agent is deployed in the cloud. Furthermore, a Local‑Privacy RAG mechanism and a heterogeneous evidence‑weighted consensus mechanism are put forward. The former prevents private knowledge from leaving the edge domain, and the latter constructs dynamic weights using evidence confidence, agent‑role reliability and cross‑agent consistency to realize traceable consensus through multi‑round divergence resolution. Validated on a privacy‑sensitive multi‑expert medical consultation scenario without domain‑specific customization, the proposed architecture is evaluated from reasoning quality, privacy exposure and system‑performance perspectives. Ablation experiments verify individual contributions of core modules. Experimental results show that GEHMAR outperforms traditional multi‑agent schemes in accuracy, evidence completeness and task success rate. It drastically reduces raw‑data transmission volume and achieves strong data isolation with acceptable latency overhead.References
[1] Li J, Cheng X, Zhao W, et al. HaluEval: A Large‑Scale Hallucination Evaluation Benchmark for Large Language Models. Proceedings of EMNLP, 2023: 6449‑6464.
[2] Cai G, Tian R, Yang L, et al. Efficient Inference for Edge Large Language Models: A Survey. Tsinghua Science and Technology, 2025.
[3] Huang L, Yu W, Ma W, et al. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Transactions on Information Systems, 2025, 43(2): 42.
[4] Chen K, Zhou X, Lin Y, et al. A Survey on Privacy Risks and Protection in Large Language Models. Journal of King Saud University Computer and Information Sciences, 2025, 37: 163.
[5] Luo J, Zhang Y, Zhang Z, et al. Cryptography‑Based Privacy‑Preserving Large Language Models: A Lifecycle Survey of Frameworks, Methods, and Future Directions. Artificial Intelligence Review, 2026, 59: 64.
[6] Thirunavukarasu A J, Ting D S J, Elangovan K, et al. Large language models in medicine. Nature Medicine, 2023, 29: 1930‑1940.
[7] Wiest I C, Ferber D, Zhu J, et al. Privacy‑preserving large language models for structured medical information retrieval. npj Digital Medicine, 2024, 7: 257.
[8] Piccialli F, Chiaro D, Qi P, et al. Federated and Edge Learning for Large Language Models. Information Fusion, 2025, 117: 102840.
[9] Aina O K, Ha S. Deep reinforcement learning for multi‑agent coordination. Artificial Life and Robotics, 2025, 31(2): 1‑11.
[10] Maktabifard A, Földes D. Cooperative driving route planning based on asymmetric multi‑agent path planning problem with limited service area constraints. Applied Soft Computing, 2025, 185: 113959.