GenAINet: Enabling Wireless Collective Intelligence via Knowledge Transfer and Reasoning

TL;DR

GenAINet tackles the gap where wireless networks must support GenAI-driven collective intelligence. It proposes a semantic-native GenAINet architecture in which distributed GenAI agents extract semantics from heterogeneous data, maintain a knowledge model, and exchange knowledge to plan and reason, reducing data traffic while enabling arbitrary task solving. The work defines three representations for knowledge (VE, KG, TE) and three collaboration levels (information, decision, intent), and demonstrates two case studies: semantic on-device telecom-domain QnA with semantic knowledge transfer, and collaborative wireless power control via reasoning across agents. Results show substantial semantic compression with minimal loss in query accuracy and improved power efficiency through cooperative reasoning, suggesting a practical path toward CI and potentially AGI in 6G networks. The paper also discusses challenges such as edge deployment, real-time KB updates, and energy/QoS guarantees for robust operation.

Abstract

Generative Artificial Intelligence (GenAI) and communication networks are expected to have groundbreaking synergies for 6G. Connecting GenAI agents via a wireless network can potentially unleash the power of Collective Intelligence (CI) and pave the way for Artificial General Intelligence (AGI). However, current wireless networks are designed as a "data pipe" and are not suited to accommodate and leverage the power of GenAI. In this paper, we propose the GenAINet framework in which distributed GenAI agents communicate knowledge (facts, experiences, and methods) to accomplish arbitrary tasks. We first propose an architecture for a single GenAI agent and then provide a network architecture integrating GenAI capabilities to manage both network protocols and applications. Building on this, we investigate effective communication and reasoning problems by proposing a semantic-native GenAINet. Specifically, GenAI agents extract semantics from heterogeneous raw data, build and maintain a knowledge model representing the semantic relationships among pieces of knowledge, which is retrieved by GenAI models for planning and reasoning. Under this paradigm, different levels of collaboration can be achieved flexibly depending on the complexity of targeted tasks. Furthermore, we conduct two case studies in which, through wireless device queries, we demonstrate that extracting, compressing and transferring common knowledge can improve query accuracy while reducing communication costs; and in the wireless power control problem, we show that distributed agents can complete general tasks independently through collaborative reasoning without predefined communication protocols. Finally, we discuss challenges and future research directions in applying Large Language Models (LLMs) in 6G networks.

Figures (8)

• Figure 1: Proposed GenAINet architecture (left) and agent architecture (right) for protocol and application management. GenAINet leverages unified GenAI agents to manage communication protocols (e.g., PHY, MAC) and applications (e.g., robotics, vehicles). Each device is associated with an agent that communicates using semantic knowledge. The GenAI agent includes four components: perception (collects information from the environment and other agents), memory (stores long/short-term knowledge in formats like vector embeddings and knowledge graphs), action (available tools), and planning (reasoning and decision-making using LLMs).
• Figure 2: Proposed pipeline of multi-modal semantic extraction, retrieval, and reasoning on GenAINet agents. Multi-modal LLMs unify information from various data modalities as semantics, enabling effective sensor data merging. Common and evolving knowledge are semantically represented (text or embeddings) to support compression, exchange, and retrieval, avoiding redundant multi-modal raw data exchange. LLMs reason and plan based on information, knowledge, and prior experiences, refining knowledge models over time. Our framework is highly adaptable to diverse tasks and heterogeneous data modalities.
• Figure 3: Three levels of collaboration in GenAINet: information, decision, and goal/intent levels. At the information level, only relevant information messages are exchanged. At the decision level, agents share decisions while considering the decisions of others. At the goal/intent level, reasoning paths are generated by considering the goals of other agents, and only the goals are shared, reducing transmission overhead. For instance, car avoidance highlights how exchanging intent information rather than raw data like images suffices. These collaboration levels are adaptable to task complexity, reflecting human-like interactions in GenAINet.
• Figure 4: Common knowledge transfer from a cloud llm to an on-device llm with semantic compression and rag. When an on-device llm receives a knowledge query beyond its capabilities, it first searches for its semantic vector database to find the corresponding knowledge item with maximum semantic similarity and leverage rag to generate accurate response. Notice that the on-device knowledge base is obtained by semantic compression of the huge cloud knowledge base.
• Figure 5: Prompt template for collaborative power control leveraging GenAINet with memory and reasoning. The target of all agents is shared in the beginning then only each agent has accessed to its local memory and leverage it reasoning capabilities.