Communications-Incentivized Collaborative Reasoning in NetGPT through Agentic Reinforcement Learning

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.

💡 Research Summary

The paper addresses a fundamental limitation of current AI deployments in next‑generation (xG) wireless networks: they are largely siloed, performing isolated optimizations without the ability to adapt, delegate tasks, or collaborate across multiple specialized agents. To overcome this, the authors propose a unified “NetGPT” framework that embeds a large language model (LLM) core within the network and equips it with the capability to either reason autonomously or invoke domain‑specific agents through structured agentic communication.

Architecture. NetGPT is organized into three logical layers: (1) a core LLM that serves as the primary service endpoint for terminals and edge applications, (2) a library of specialized agents distributed across the user‑plane, compute‑plane, and control‑plane, and (3) a shared infrastructure providing knowledge bases, registries, and persistent memory. Each agent declares its capabilities, performance metrics (throughput, latency, cost), and supported “action types” (e.g., NetworkAnalysis, DataRetrieval). An “agent card” abstracts these declarations, allowing the core to discover, select, and orchestrate agents regardless of the underlying communication protocol (A2A, ACP, ANP).

Collaborative Reasoning Workflow. When a request arrives, the core first enriches and interprets the user intent. It then evaluates task difficulty, time‑sensitivity, and required knowledge. If the task can be solved locally, the core proceeds with internal reasoning; otherwise it decomposes the task into sub‑tasks aligned with predefined action types. Candidate agents are retrieved via their cards, and a routing decision is made using one of three mechanisms: rule‑based (deterministic, high‑reliability), machine‑learning‑based (trained on historic intent‑routing pairs), or LLM‑based (leveraging contextual reasoning). The selected agent is invoked through a structured JSON‑like request, its progress is monitored, and the returned results are integrated into a coherent final answer.

Agentic Reinforcement Learning (Agentic RL). To enable the system to learn when to delegate and which agents to pick, the authors extend the conventional LLM training pipeline (Supervised Fine‑Tuning → RLHF → RLVR) to a partially observable Markov decision process (POMDP) where actions include both token generation and external agent calls. The training pipeline introduces three key innovations:

1. Dynamic Masked Loss – The loss is amplified for steps where the external agent’s output is uncertain or erroneous, encouraging the core to model external uncertainty explicitly.
2. Entropy‑Guided Exploration – During high‑uncertainty decision points, the policy’s token‑selection entropy is increased to explore alternative delegation strategies; in low‑uncertainty phases entropy is penalized to promote convergence.
3. Multi‑Objective Reward – The reward function aggregates (i) task accuracy, (ii) latency, and (iii) resource consumption (compute and communication cost). By weighting these objectives, the policy learns to balance quality against efficiency, a crucial trade‑off in real‑time network operations.

Continuous Evolution. After each episode, performance metrics (success rate, latency, cost) and user QoE feedback are fed back as reward signals, allowing the policy network to be updated continually. This creates a self‑improving loop that adapts to changing network topologies, new agents, and fluctuating resource availability without manual re‑engineering.

Experimental Validation. The authors evaluate the framework on a network root‑cause analysis scenario. A user reports poor signal quality at a specific location and time. NetGPT interprets the request, determines that a “NetworkAnalysis” action is required, selects an analysis agent, and automatically retrieves drive‑test logs and cell parameters. The agent returns processed data, which NetGPT integrates into a diagnosis stating that an excessive downtilt angle caused coverage loss. Compared with a baseline single‑LLM system, the collaborative approach reduces end‑to‑end latency by over 30 %, improves diagnostic accuracy by roughly 15 %, and cuts overall compute/communication cost by about 20 %.

Contributions and Impact. The paper makes three major contributions: (1) a modular, protocol‑agnostic architecture that cleanly separates core reasoning from domain‑specific execution, (2) a novel Agentic RL formulation tailored to partially observable, stochastic environments with structured external actions, and (3) a practical training pipeline that combines masked loss, entropy‑driven exploration, and multi‑objective rewards to achieve stable learning in complex multi‑agent settings. By demonstrating measurable gains in a realistic telecom use‑case, the work provides a concrete roadmap for deploying AI‑native, self‑evolving intelligence across future xG networks, paving the way for more resilient, scalable, and cost‑effective network management.