Communications-Incentivized Collaborative Reasoning in NetGPT through Agentic Reinforcement Learning

TL;DR

The paper tackles the need for AI-native, collaborative reasoning in next-generation ($xG$) networks by proposing NetGPT, a unified framework where a core can reason autonomously or delegate to domain agents through agentic communication. It introduces a modular architecture with an agent-governed registry, intent evaluation, adaptive orchestration, and continuous evolution, reinforced by an agentic reinforcement learning training pipeline that combines routing, masked loss, entropy-guided exploration, and multi-objective rewards. A two-phase training strategy—supervised fine-tuning followed by agentic RL—enables stable, multi-turn collaborative reasoning in partially observable, stochastic environments, demonstrated via a proof-of-concept with network analysis and protocol query actions. The results suggest NetGPT can learn when to invoke external agents, how to balance accuracy and resource costs, and how to integrate multi-agent outputs into cohesive network decisions, moving toward autonomous sensing, reasoning, and action in complex telecom settings. This work lays a foundation for self-evolving, cognition-driven $xG$ networks capable of scalable, trustworthy coordination across distributed planes and tools.

Abstract

The evolution of next-Generation (xG) wireless networks marks a paradigm shift from connectivity-centric architectures to Artificial Intelligence (AI)-native designs that tightly integrate data, computing, and communication. Yet existing AI deployments in communication systems remain largely siloed, offering isolated optimizations without intrinsic adaptability, dynamic task delegation, or multi-agent collaboration. In this work, we propose a unified agentic NetGPT framework for AI-native xG networks, wherein a NetGPT core can either perform autonomous reasoning or delegate sub-tasks to domain-specialized agents via agentic communication. The framework establishes clear modular responsibilities and interoperable workflows, enabling scalable, distributed intelligence across the network. To support continual refinement of collaborative reasoning strategies, the framework is further enhanced through Agentic reinforcement learning under partially observable conditions and stochastic external states. The training pipeline incorporates masked loss against external agent uncertainty, entropy-guided exploration, and multi-objective rewards that jointly capture task quality, coordination efficiency, and resource constraints. Through this process, NetGPT learns when and how to collaborate, effectively balancing internal reasoning with agent invocation. Overall, this work provides a foundational architecture and training methodology for self-evolving, AI-native xG networks capable of autonomous sensing, reasoning, and action in complex communication environments.

Figures (6)

• Figure 1: Communications-incentivized collaborative reasoning in NetGPT.
• Figure 2: (Top): NetGPT core-based collaborative reasoning workflow from intent interpretation, action-based orchestration, and execution. (Bottom): Collaborative reasoning example for network root cause analysis.
• Figure 3: Development of training phase for LLM agents.
• Figure 4: Training pipeline for NetGPT.
• Figure 5: Examples of failure modes and the eventual convergence of agentic RL.