PaperlayerView exact revision on arXiv

Source-linked AI summary

Toward Edge General Intelligence with Multiple-Large Language Model (Multi-LLM): Architecture, Trust, and Orchestration

Haoxiang Luo, Yinqiu Liu, Ruichen Zhang, Jiacheng Wang, Gang Sun, Dusit Niyato, Hongfang Yu, Zehui Xiong, Xianbin Wang, Xuemin Shen

arXiv:2507.00672v1cs.NIcs.DC

TL;DR

Edge environments require adaptable intelligence despite resource, reliability, privacy, and trust constraints. This survey examines multi-LLM architectures, collaboration mechanisms, deployment technologies, and governance for edge general intelligence, concluding that lightweight, multimodal, trustworthy systems remain central future directions.

  • Problem

    Existing edge AI and single-LLM systems face resource, hallucination, reliability, privacy, and trust challenges, while comprehensive knowledge of multi-LLM collaboration for edge general intelligence remains limited.

  • Method

    The survey synthesizes multi-LLM architectures, collaboration modes, ensemble integration, compression, multimodal processing, deployment strategies, and trustworthy governance for edge computing.

  • Results

    The survey identifies multi-LLM systems as a promising approach for reliable edge intelligence while documenting resource constraints, collaboration risks, and the need for trusted governance.

  • Takeaways & Limitations

    Future progress depends on lightweight multimodal models, cross-domain knowledge transfer, dynamic adaptation, transparent governance, and stronger reasoning methods.

Abstract

from arXiv · show

Edge computing enables real-time data processing closer to its source, thus improving the latency and performance of edge-enabled AI applications. However, traditional AI models often fall short when dealing with complex, dynamic tasks that require advanced reasoning and multimodal data processing. This survey explores the integration of multi-LLMs (Large Language Models) to address this in edge computing, where multiple specialized LLMs collaborate to enhance task performance and adaptability in resource-constrained environments. We review the transition from conventional edge AI models to single LLM deployment and, ultimately, to multi-LLM systems. The survey discusses enabling technologies such as dynamic orchestration, resource scheduling, and cross-domain knowledge transfer that are key for multi-LLM implementation. A central focus is on trusted multi-LLM systems, ensuring robust decision-making in environments where reliability and privacy are crucial. We also present multimodal multi-LLM architectures, where multiple LLMs specialize in handling different data modalities, such as text, images, and audio, by integrating their outputs for comprehensive analysis. Finally, we highlight future directions, including improving resource efficiency, trustworthy governance multi-LLM systems, while addressing privacy, trust, and robustness concerns. This survey provides a valuable reference for researchers and practitioners aiming to leverage multi-LLM systems in edge computing applications.

I. INTRODUCTION

Edge computing brings intelligence closer to data sources to reduce latency and bandwidth use, while EGI extends edge AI through generalist LLM capabilities. This survey addresses the gap in understanding how multiple LLMs can collaborate in edge networks to achieve EGI.

  • Edge computing processes data near sources such as sensors, cameras, vehicles, and drones, reducing latency and bandwidth use for time-sensitive applications.
  • LLMs support EGI by providing generalist understanding, generation, and reasoning over language and multimodal data without task-specific retraining.
  • Individual LLMs may struggle with heterogeneous contexts, outdated knowledge, hallucinations, and differing biases, motivating collaboration among specialized models.
  • The survey reviews multi-LLM architectures, deployment strategies, enabling technologies, and applications across edge environments, with attention to cloud–edge differences.
  • Existing work has studied edge LLM deployment and multi-LLM collaboration separately, but lacks a comprehensive investigation of their combination for EGI.

C. Our Contributions

The survey reviews multi-LLM architectures, deployment strategies, enabling technologies, applications, trust mechanisms, datasets, and future challenges for edge general intelligence.

  • The survey reviews architectural designs and deployment strategies for multi-LLM systems operating in edge computing environments.
  • It summarizes four edge applications—elderly care, smart grid inspection, intelligent transportation, and LAENets—to illustrate multi-LLM potential, especially in mobile scenarios.
  • The survey categorizes enabling technologies spanning model compression, resource orchestration, model context protocol, privacy protection, fine-tuning, and multimodal information fusion.
  • It analyzes trustworthy multi-LLM mechanisms and provides a case study and open-source dataset summary for edge deployment research.
  • Future systems should use lightweight multimodal models, improve cross-domain generalization and adaptation, establish trustworthy governance, and strengthen reasoning methods.

1) Cooperative Collaboration:

Multi-LLM systems use cooperation, adversarial interaction, and ensemble integration to combine model strengths, validate outputs, and improve robustness, with different interaction costs.

  • Cooperative Collaboration: Cooperative collaboration divides complex tasks, shares information, and iteratively refines outputs by assigning complementary roles to multiple LLMs.
  • Cooperative Collaboration: Cooperative multi-LLM systems support fairness, bias mitigation, contextual reasoning, and multimodal fusion through complementary knowledge and cross-verification.
  • Competitive or Adversarial Collaboration: Adversarial collaboration has models challenge and critique one another’s outputs to stress-test correctness and consistency through debate.
  • Competitive or Adversarial Collaboration: Competitive interaction creates checks and balances that promote robustness but require more complex interaction protocols.
  • Ensemble Integration: Ensemble methods run multiple LLMs independently and combine their outputs, commonly through voting or confidence-weighted aggregation.
  • Ensemble Integration: Ensemble approaches benefit from model diversity and typically produce stable, reliable outputs without intensive inference-time interaction.

B. Differences among Traditional AI, Single LLM, and Multi-LLM

Traditional AI models are narrow and task-specific, while single LLMs provide broader reasoning but remain resource-intensive and vulnerable to hallucination. Multi-LLM systems combine specialized models to expand task coverage, support flexible scaling, improve robustness, and address privacy and reliability concerns through collaboration.

  • Traditional AI: Traditional AI models are optimized for defined tasks but have limited reasoning and generalization beyond their training distributions.Adding new functionality typically requires a separate model or algorithm.
  • Single LLM: Single LLMs can handle diverse tasks through prompting and few-shot examples, but edge deployment requires substantial memory and compute.Even compressed or smaller models may require about 500 MB of RAM and advanced processors.
  • Single LLM: Single LLMs may produce biased, false, outdated, or illusory content because of incomplete training data and architectural limitations.These limitations make the single-model approach a potential reliability bottleneck in dynamic edge settings.
  • Multi-LLM: Multi-LLM systems use specialized models and collaboration to cover broader or more complex task spaces, including tasks requiring legal interpretation and numerical reasoning.Different knowledge bases and model strengths support complementary expertise.
  • Multi-LLM: Multi-LLM systems allow models to be added or replaced without retraining the entire system and distribute subtasks for greater fault tolerance.Cross-verification and ensemble effects can produce more accurate and robust results.
  • Multi-LLM: Multi-LLM systems can route sensitive data to private models, while collaboration helps reduce hallucinations and bias across complex scenarios.Their resource footprint, coordination requirements, and susceptibility to malfunctioning or malicious models create edge deployment and trust challenges.

C. Differences of Multi-LLM in the Cloud and Edge

Multi-LLM edge deployments process data closer to its source, reducing latency and supporting local handling, but they face tighter compute, memory, and power constraints than cloud systems. These constraints favor distributed inference and motivate specialized collaboration for privacy-sensitive, multimodal applications such as elderly care.

  • Edge versus Cloud: Edge inference generally reduces latency by processing data near its source, whereas cloud inference adds round-trip delays to distant data centers.The edge can also support local handling for real-time and privacy-sensitive applications.
  • Edge versus Cloud: Edge environments have limited compute, memory, and power; a 7 B-parameter LLaMA 2 requires 28 GB RAM without compression or partitioning.Cloud data centers can use abundant GPUs and memory more easily.
  • Edge versus Cloud: Because of edge constraints, multi-LLM inference may be split across multiple edge nodes or coordinated between devices and edge servers.This distributed architecture contrasts with cloud deployments in which the ensemble runs centrally.
  • Applications: The survey identifies elderly care, smart grid inspection, intelligent transportation, and LAENets as representative edge applications for multi-LLM systems.The latter three are described as especially promising for mobile scenarios.
  • Elderly Care: Multi-LLM systems support elderly care by assigning specialized models to medical advice, home automation, conversation, and other functions.A reported system combined GPT-3, GPT-4, and six other LLMs for payments, daily assistance, safety monitoring, and emotional support.
  • Elderly Care: On-device multi-LLM processing keeps sensitive personal and medical data local while reducing latency for time-critical support such as fall detection and medication reminders.Access control and privacy-preserving collaboration help protect user data during model cooperation.

B. Smart Grid Inspection

Multi-LLM systems support smart grid inspection by coordinating specialized models across heterogeneous sensor and imagery data, while edge deployment enables real-time, resilient responses.

  • Smart Grid Inspection: Multiple LLMs can jointly analyze IoT sensor streams and drone imagery of power lines for holistic grid inspection.Specialized models can focus on particular modalities or sub-problems before combining their results.
  • Smart Grid Inspection: Deploying LLMs at substations, control centers, and inspection drones minimizes communication delays and supports local operation when cloud connectivity fails.The arrangement reduces fault-response times and bandwidth usage.
  • Smart Grid Inspection: Multimodal data fusion combines sensor inputs with other information to support grid analysis across distributed edge locations.Distributed placement is motivated by the grid’s geographic spread and real-time response requirements.

E. Lessons Learned

Efficient edge multi-LLM deployment combines model compression with adaptive routing, cooperative scheduling, and lightweight inference to balance resource constraints, latency, and accuracy.

  • Lessons Learned: 40% fewer parameters preserve 97% of BERT’s performance, while other compressed models reduce size and inference cost for edge deployment.TinyBERT achieves 96.8% of BERT-base accuracy at 7.5× smaller and 9.4× faster, and MobileBERT is 4.3× smaller and 5.5× faster.
  • Lessons Learned: Query routing and cascaded inference send easy requests to compact models and escalate difficult cases to larger models.Confidence-based gating or ensemble agreement supports distributed decisions while reducing average latency and energy.
  • Lessons Learned: Cloud–edge co-training uses split learning and LoRA-style adapters so edge devices exchange intermediate representations or lightweight updates instead of full models.Parallel split federated learning enables large-model fine-tuning on edge GPUs with reduced communication burden.
  • Lessons Learned: Lightweight inference uses adapters, token filtering, early exits, and cloud–edge pipelines to meet strict latency and energy targets.Early exits terminate computation when confidence is high, while uncertain processing can be offloaded.
  • Lessons Learned: Distributed orchestration assigns model layers, inference tasks, and device loads through cooperative scheduling, dynamic programming, and multi-agent reinforcement learning.These methods target lower latency, higher throughput, and balanced accuracy under time and energy constraints.

C. Model Context Protocol

Edge multi-LLM systems require protocols for distributing computation, managing shared context, and protecting private information during collaboration.

  • Model Context Protocol: Model context management compresses or streams inputs to fit edge constraints, with ICAE achieving roughly 4× context compression.The approach uses a lightweight autoencoder to summarize long contexts into a small set of memory slots.
  • Model Context Protocol: Model-distributed inference partitions one generation task across devices whose layers pipeline computation to reduce idle time.MDI-LLM exemplifies this strategy.
  • Model Context Protocol: Private and shared memory banks preserve user-specific interactions while selectively sharing reusable knowledge across LLMs.Selective sharing avoids unnecessary duplication under limited edge resources.
  • Model Context Protocol: Retrieval-based memory fusion lets agents retrieve relevant prompt–response examples from a communal store and add them to new in-context prompts.This enables reuse of other agents’ experiences during inference.
  • Model Context Protocol: Peer-to-peer prompt exchange broadcasts queries to neighboring LLMs, after which one model aggregates proposals into a final answer.The Mixture-of-Agents system uses intermittent device communication without a central server.
  • Model Context Protocol: Privacy protection combines local inference, federated updates, cryptographic safeguards, differential privacy, and role-based filtering of cross-LLM information.These measures limit raw-data exposure while addressing leakage from updates, prompts, and model outputs.

E. LLM Fine-Tuning

Edge LLM fine-tuning reduces adaptation costs through parameter-efficient, federated, split, and transfer-learning strategies, while multimodal systems align specialized model outputs.

  • LLM Fine-Tuning: LoRA freezes pretrained LLM weights and adds trainable low-rank matrices to adapt models to specific edge tasks.This supports fine-tuning for applications such as smart-home voice commands.
  • LLM Fine-Tuning: Federated fine-tuning lets edge devices update local LLM copies and share model updates without directly exchanging raw data.Differential privacy and secure multi-party computation can protect shared updates.
  • LLM Fine-Tuning: Split learning reduces client computation and communication by exchanging activations and gradients instead of the entire LLM.SplitLoRA applies this approach across clients and servers.
  • LLM Fine-Tuning: Transfer learning reuses knowledge from a general edge task to reduce data and computation needed for related downstream fine-tuning.A model trained for general natural language processing can be adapted to specific edge applications.
  • LLM Fine-Tuning: Multimodal multi-LLM systems assign modalities to specialized models and fuse or cross-validate their outputs for comprehensive scene understanding.Visual, textual, and other modality-specific results can be combined for more robust decisions.
  • LLM Fine-Tuning: Attention-based fusion and cross-modal feature alignment connect heterogeneous modality representations into a shared semantic space.These methods focus on important information and improve collaboration across LLMs.

G. Lessons Learned

Multi-LLM edge systems must balance performance, resource use, and security while addressing trustworthiness, malicious behavior, coordination failures, and limited transparency. The survey emphasizes trustworthy collaboration and supports further work through datasets.

  • G. Lessons Learned: Edge multi-LLM systems face a core performance-resource-security constraint.Multimodal fusion and context protocols improve performance, while lightweight models and resource orchestration address resource limits; privacy and fine-tuning support reliability and accuracy.
  • G. Lessons Learned: Trustworthiness requires authentic and reliable responses that attackers cannot maliciously tamper with.Open wireless environments, scattered devices, and complex edge conditions weaken response trustworthiness.
  • G. Lessons Learned: Different model outputs make it difficult to determine which response has the best credibility and quality.The outputs differ because LLMs use different corpora, training methods, and scenario orientations.
  • G. Lessons Learned: Malicious devices, operators, or purpose-built models can produce misleading responses and expose privacy or economic interests.The survey cites viruses, Trojans, operator intent, and malicious models such as WormGPT as threats.
  • G. Lessons Learned: Centralized coordination introduces single-point failure, efficiency bottlenecks, and risks from unverifiable or malicious LLM responses.A compromised or failed central node can disrupt collaboration, while malicious models may contaminate collective results.
  • G. Lessons Learned: Opaque collaboration makes response sources difficult to trace and reliability difficult to audit.Improved response quality does not by itself provide provenance for generated content.

1) Blockchain-driven multi-LLM:

Blockchain-driven multi-LLM collaboration uses decentralized consensus, verification, and traceable storage to select and preserve responses. The surveyed case study reports higher user ratings for trustworthy WBFT-based collaboration than for single-LLM and non-blockchain multi-LLM systems.

  • 1) Blockchain-driven multi-LLM:: Blockchain consensus reduces reliance on trusted third parties while improving response traceability and credibility.Consensus addresses centralized coordination bottlenecks and single-point failure, while blockchain immutability preserves response provenance.
  • 1) Blockchain-driven multi-LLM:: The workflow comprises user initiation, distributed LLM response creation, voting-based consensus, block formation, and decentralized blockchain extension.Responses are broadcast through the blockchain P2P network, and the selected result is stored with a hash and timestamp.
  • 1) Blockchain-driven multi-LLM:: WBFT assigns voting rights jointly according to response quality and LLM credibility.The consensus weights these two factors through parameters α and β.
  • 1) Blockchain-driven multi-LLM:: 15 volunteers rated individual LLMs, non-blockchain multi-LLM, and trustworthy WBFT-driven multi-LLM under different α and β settings.The comparison used user ratings of generation capabilities across the three schemes.
  • 1) Blockchain-driven multi-LLM:: The trustworthy WBFT-driven system scored significantly above both a single LLM and multi-LLM without blockchain participation.The results also indicate that response quality affects user satisfaction and LLM credibility affects WBFT performance.
  • 1) Blockchain-driven multi-LLM:: Blockchain-integrated multi-LLM is presented for elderly care, smart-grid inspection, intelligent transportation, and LAENets.These scenarios use collaborative verification, consensus, encryption, or on-chain records to support safer decisions and traceability.

B. Open-Source Datasets

The survey catalogs publicly available datasets that support multi-LLM training, evaluation, preference optimization, conversation, and domain-specific fine-tuning. It also notes that edge applications of multi-LLM remain in an early stage.

  • B. Open-Source Datasets: The dataset survey identifies public resources for multi-LLM training and performance evaluation.The section introduces the datasets and summarizes their roles in supporting EGI research.
  • B. Open-Source Datasets: FSPO uses a small amount of real user preference data to align LLM outputs with individual preferences.Its purpose is preference optimization rather than broad multi-task evaluation.
  • B. Open-Source Datasets: BabbleBeaver provides prompt examples for conversations among multiple LLMs, including OpenAI, Gemini, Mistral, Anthropic, Cohere, and Ollama.The resource supports cross-model conversational interaction.
  • B. Open-Source Datasets: MMLU measures multi-domain knowledge understanding and problem-solving across a large-scale multitask test set.It is intended to assess performance across a wide range of knowledge domains.
  • B. Open-Source Datasets: LLM-QA compares multiple LLMs on multiple-choice tasks to identify the best-performing model for a specific task.Its evaluation focus is comparative task accuracy.
  • B. Open-Source Datasets: GSM8K supports training and evaluation of primary-school mathematics solving, including repeated self-error identification and correction.The dataset is used to assess whether models can retry until reaching a correct solution.
  • B. Open-Source Datasets: ChatGLM provides history-related prompts for fine-tuning a Chinese chatbot.The prompts are provided in JSON format.
  • B. Open-Source Datasets: Multi-LLM application in edge computing remains in its early stages despite progress in compression, orchestration, privacy, fine-tuning, and multimodal fusion.The survey frames future directions as necessary for advancing multi-LLM-enabled EGI.

A. Lightweight Multimodal LLM

Future edge-oriented multi-LLM systems need lightweight multimodal architectures, cross-domain adaptation, trustworthy governance, stronger security, and advanced reasoning methods. The survey presents these directions as central to efficient and reliable EGI.

  • A. Lightweight Multimodal LLM: Lightweight multimodal models are needed to process diverse data under edge resource constraints.The survey cites a multimodal emotion-recognition architecture using approximately 2.7 M parameters and recommends compression, quantization, pruning, and knowledge distillation.
  • A. Lightweight Multimodal LLM: Efficient multimodal architecture can use depth-separable convolution, hardware adaptation, and acceleration to reduce redundancy and improve computational efficiency.The proposed direction draws on lightweight network designs such as MobileNet.
  • A. Lightweight Multimodal LLM: Cross-domain generalization remains a key challenge for adapting multi-LLM systems to new fields.Adversarial training can reduce source-target feature differences, while multi-domain pre-training and transfer learning support broader adaptation.
  • A. Lightweight Multimodal LLM: Credibility is identified as essential for the wide application of multi-LLM systems.Future work includes explanation generators and attention visualization to expose model concerns and reasoning paths for multimodal outputs.
  • A. Lightweight Multimodal LLM: Trusted AI governance requires stronger security, robustness, vulnerability repair, and privacy-protection mechanisms.Suggested techniques include adversarial training, data augmentation, vulnerability detection, and advanced cryptography.
  • A. Lightweight Multimodal LLM: RAG, RAP, Chain of Thought, world models, and agentic AI are presented as reasoning approaches for complex multimodal and collaborative tasks.These methods support retrieval, stepwise decomposition, cognitive consistency, task collaboration, and autonomous interaction.
  • A. Lightweight Multimodal LLM: The survey concludes that multi-LLM systems have potential to advance edge computing toward more intelligent and autonomous applications.This conclusion is tied to lightweight architecture, trusted governance, robust decisions, reliable responses, and privacy-aware deployment.
Loading 2507.00672v1…