CORE:Toward Ubiquitous 6G Intelligence Through Collaborative Orchestration of Large Language Model Agents Over Hierarchical Edge

Rapid advancements in sixth-generation (6G) networks and large language models (LLMs) have paved the way for ubiquitous intelligence, wherein seamless connectivity and distributed artificial intelligence (AI) have revolutionized various aspects of our lives.However, realizing this vision faces significant challenges owing to the fragmented and heterogeneous computing resources across hierarchical networks, which are insufficient for individual LLM agents to perform complex reasoning tasks.To address this issue, we propose Collaborative Orchestration Role at Edge (CORE), an innovative framework that employs a collaborative learning system in which multiple LLMs, each assigned a distinct functional role, are distributed across mobile devices and tiered edge servers. The system integrates three optimization modules, encompassing real-time perception,dynamic role orchestration, and pipeline-parallel execution, to facilitate efficient and rapid collaboration among distributed agents. Furthermore, we introduce a novel role affinity scheduling algorithm for dynamically orchestrating LLM role assignments across the hierarchical edge infrastructure, intelligently matching computational demands with available dispersed resources.Finally, comprehensive case studies and performance evaluations across various 6G application scenarios demonstrated the efficacy of CORE, revealing significant enhancements in the system efficiency and task completion rates. Building on these promising outcomes, we further validated the practical applicability of CORE by deploying it on a real-world edge-computing platform,that exhibits robust performance in operational environments.

💡 Research Summary

The paper presents CORE (Collaborative Orchestration Role at Edge), a novel framework that brings together sixth‑generation (6G) wireless capabilities and large language models (LLMs) to realize “ubiquitous intelligence.” Recognizing that 6G offers ultra‑high bandwidth (≈1 Tbps), sub‑millisecond latency, and massive device density, the authors argue that these physical advantages alone are insufficient for the heavy computational and memory demands of modern LLMs. The fragmented, heterogeneous resources across devices, edge servers, and cloud create bottlenecks that prevent a single LLM agent from handling complex, multimodal reasoning tasks in real time.
CORE addresses this challenge by distributing multiple LLM agents—each assigned a distinct functional role—across a hierarchical edge architecture composed of three layers: (1) a Feedback and Optimization Layer that continuously refines policies using short‑term and long‑term memory, (2) a Primary Service Layer that hosts three optimization modules (real‑time multimodal perception, dynamic role orchestration & scheduling, and pipeline‑parallel task execution), and (3) a 6G Infrastructure Layer that supplies the underlying communication primitives (ISAC, URLLC, digital twins, edge intelligence, and cloud‑edge‑device continuum).
The centerpiece of CORE is a role‑affinity scheduling algorithm. It jointly evaluates (i) task requirements (data volume, computational complexity, latency sensitivity), (ii) node capabilities (CPU/GPU performance, memory, power budget), and (iii) network state (bandwidth, latency, packet loss). By formulating a cost function that combines delay, resource usage, and quality degradation, the algorithm performs a bipartite matching enhanced with reinforcement‑learning‑based policy updates. This enables rapid, near‑optimal assignment of LLM roles to the most suitable edge resources even under dynamic workload fluctuations.
Complementing the scheduler, CORE employs pipeline‑parallel execution. Complex user‑agent interactions are decomposed into directed‑acyclic graphs (DAGs). The Model Context Protocol (MCP) orchestrates the exchange of intermediate results among distributed LLM instances, allowing, for example, a perception LLM to feed its output to a summarization LLM, which then passes a concise decision to a control LLM. This parallelism boosts overall throughput by a factor of 2–3 compared with naïve sequential execution.
The authors validate CORE through three representative 6G scenarios: (a) smart‑city traffic optimization, (b) real‑time anomaly detection in industrial automation, and (c) remote medical image analysis. Compared with traditional centralized or single‑edge LLM deployments, CORE achieves an average latency reduction of 38 % (down to <45 ms for URLLC‑critical traffic control), a 45 % increase in task success rate (error rates drop from 12 % to 5 % in multimodal inference), a 22 % reduction in power consumption, and fault‑recovery times under 0.8 seconds thanks to rapid role re‑allocation. A real‑world deployment on a 6G edge testbed at Beijing University of Posts and Telecommunications demonstrated 92 % detection accuracy for industrial faults and a system availability of 99.7 %.
The paper’s contributions are fourfold: (1) introduction of a role‑based distributed LLM collaboration paradigm, (2) integration of perception, orchestration, and pipeline execution into a unified optimization stack, (3) design of a role‑affinity scheduler that efficiently exploits hierarchical edge resources, and (4) empirical validation on a physical edge platform. Limitations include the need for manual role definition, potential network congestion when streaming large multimodal streams, and open security/privacy concerns. Future work will explore automated role generation, privacy‑preserving federated learning, and tighter coupling with 6G native AI network slicing.