NextG-GPT: AI-Native Wireless Networks

• NextG-GPT is an umbrella framework for AI-native wireless networks that combines large language models, retrieval-augmented generation, and cloud–edge orchestration to shift from connectivity-focused operations.
• The architecture features modular components such as semantic search using FAISS, dual-stage prompt processing, and personalized inference across edge and cloud platforms for dynamic network control.
• Recent advancements incorporate agentic collaboration and reinforcement learning, optimizing task delegation, resource usage, and communication accuracy in complex wireless environments.

Searching arXiv for papers on NextG-GPT / NetGPT to ground the article in current literature. NextG-GPT denotes a line of AI-native wireless-network frameworks that integrate LLMs, retrieval-augmented generation, cloud–edge orchestration, and, in later formulations, agentic collaboration for next-generation and beyond-next-generation communication systems. Across the literature, the term appears in closely related forms including “NextG-GPT” and “NetGPT,” with a shared objective: to move wireless networking from connectivity-centric operation toward architectures in which reasoning, knowledge retrieval, model adaptation, and network control are embedded into the communication fabric itself (Nazar et al., 25 May 2025, Chen et al., 2023, Tong et al., 2023). The concept spans multiple layers of abstraction, from a retrieval-augmented question-answering system for wireless research support to a foundation-model paradigm for wireless communications, a cloud–edge personalized generative-service architecture, and an agentic framework in which a NetGPT core delegates sub-tasks to specialized agents under reinforcement learning (Nazar et al., 25 May 2025, Yu et al., 31 Jan 2026).

1. Origins, nomenclature, and research trajectory

The earliest formulation in this literature is NetGPT, defined as “the foundation models for wireless communications,” with emphasis on generality, performance gain, management, and collaboration in future wireless networks (Tong et al., 2023). In that framing, NetGPT is not merely an application-layer assistant but a foundation-model stack spanning data ingestion, pre-training, fine-tuning, inference, lifecycle management, and cross-site updating. The terminology therefore places wireless-domain tokenization, transformer pre-training, and constrained fine-tuning at the center of network intelligence (Tong et al., 2023).

A second line of work presents NetGPT as “a Native-AI Network Architecture Beyond Provisioning Personalized Generative Services,” where the core architectural problem is collaborative cloud–edge orchestration of heterogeneous LLMs (Chen et al., 2023). In this view, NetGPT is an AI-native network architecture that uses small edge LLMs and larger cloud LLMs in a dual-stage prompt workflow, with explicit support from a new Computing Plane (CmP) and logical AI workflow mapping (Chen et al., 2023).

The designation NextG-GPT is used in a later framework that integrates retrieval-augmented generation (RAG) and open-source LLMs with a domain-specific knowledge base for wireless systems and communication research (Nazar et al., 25 May 2025). Here, the system is oriented toward context-aware real-time support for researchers and network operators, with a modular retrieval, fusion, and generation pipeline backed by telecom corpora such as ORAN-13K-Bench, TeleQnA, TSpec-LLM, and Spec5G (Nazar et al., 25 May 2025).

The most recent formulation extends the line into agentic collaboration. “Communications-Incentivized Collaborative Reasoning in NetGPT through Agentic Reinforcement Learning” introduces a unified agentic NetGPT framework in which a NetGPT core can either reason autonomously or delegate sub-tasks to domain-specialized agents via agentic communication (Yu et al., 31 Jan 2026). This suggests an evolution from knowledge-grounded answering and cloud–edge personalization toward distributed reasoning and closed-loop action under partial observability.

2. Architectural patterns across the literature

Despite differing emphases, the major papers converge on a modular architecture. In the RAG-oriented NextG-GPT system, the overall pipeline combines three stages: retrieval through semantic search over a domain-specific vector store, fusion or prompt construction from top-ranked context passages plus the query, and generation by an LLM conditioned on the fused prompt (Nazar et al., 25 May 2025). The preprocessing path is explicit: data extraction, chunking, embedding, index build, semantic search and ranking, then LLM inference with retrieved context (Nazar et al., 25 May 2025).

In the cloud–edge NetGPT architecture, the deployment is explicitly heterogeneous. The edge tier uses GPT-2-base, approximately $1.2 \times 10^8$ parameters, in macro or pico base stations or MEC servers, while the cloud tier uses LLaMA-7B, approximately $6.7 \times 10^9$ parameters, on large GPU clusters (Chen et al., 2023). A “concise prompt” $P_{\mathrm{con}}$ is sent to the edge; the edge LLM expands it into a “comprehensive prompt” $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$; and, depending on the edge confidence score relative to $\rho$, the edge either responds locally or forwards $P_{\mathrm{com}}$ to the cloud, where the cloud LLM generates the final reply $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$ (Chen et al., 2023).

The foundation-model blueprint in “Ten issues of NetGPT” specifies a broader lifecycle architecture. Data sources include radio measurements such as CSI tensors, SINR maps, and PHY waveforms; control logs such as scheduling decisions and handover records; and contextual data such as user mobility and network topology (Tong et al., 2023). These flow through data ingestion and pre-processing into a pre-training cluster hosting a Foundation Transformer (NetGPT-L0), then a model zoo and fine-tuning service yielding NetGPT-L1 and NetGPT-L2 models, with cloud and edge inference engines connected by a collaboration bus for L2-to-L1/L0 offload when local confidence is below threshold (Tong et al., 2023).

The agentic NetGPT architecture adds a core-and-agent decomposition. The NetGPT Core is a large-language-model-based “brains” of the system, with Llama-3-8B-Instruct given as the prototype example, exposing a single endpoint to edge or terminal clients for simple Q&A and complex multi-agent reasoning (Yu et al., 31 Jan 2026). Domain-specialized agents are registered through protocol-agnostic “agent cards” containing supported actions, endpoint URI, latency, accuracy, load, and cost, and these agents inhabit the User Plane, Compute Plane, and Control Plane (Yu et al., 31 Jan 2026). Toolboxes are lightweight libraries accessed through MCP, while shared infrastructure includes knowledge bases, registries, and data lakes (Yu et al., 31 Jan 2026).

A useful comparison is summarized below.

Formulation	Primary architectural emphasis	Core components named in the papers
NextG-GPT (Nazar et al., 25 May 2025)	RAG for wireless-domain support	retrieval, fusion/prompt-construction, generation, FAISS, GTE, knowledge base
NetGPT (Chen et al., 2023)	Cloud–edge personalized generative workflow	edge GPT-2-base, cloud LLaMA-7B, CmP, concise/comprehensive prompts
NetGPT (Tong et al., 2023)	Foundation-model stack for wireless communications	NetGPT-L0/L1/L2, model zoo, fine-tuning service, collaboration bus
Agentic NetGPT (Yu et al., 31 Jan 2026)	Multi-agent delegation and RL	NetGPT Core, domain-specialized agents, toolboxes, registry, shared infrastructure

These formulations are not identical systems. A plausible implication is that “NextG-GPT” functions as an umbrella label for several progressively richer architectures rather than a single frozen implementation.

3. Retrieval-augmented generation and wireless-domain knowledge grounding

In the RAG-centric NextG-GPT framework, the retrieval path begins by cleaning, segmenting, and normalizing each document into fixed-size chunks with $C = 800$ characters and overlap $C_0 = 80$ (Nazar et al., 25 May 2025). A pre-trained General Text Embedding model with context length 8192 converts each chunk into a fixed-length vector, and embeddings are stored in FAISS with hierarchical clustering for sub-linear search (Nazar et al., 25 May 2025). At query time, the system embeds the question, performs FAISS similarity search, ranks results by cosine similarity, retains the top $p = 95$th percentile, and concatenates the retained passages with the original user query (Nazar et al., 25 May 2025).

The similarity score is defined as

$6.7 \times 10^9$0

and the final prompt is described as $6.7 \times 10^9$1 (Nazar et al., 25 May 2025). Generation then uses the chosen open-source LLM with top-$6.7 \times 10^9$2 sampling at $6.7 \times 10^9$3 (Nazar et al., 25 May 2025).

The knowledge base is domain-specific and explicitly heterogeneous. The integrated datasets are ARA Documentation & APIs, ORAN-Bench-13K, TeleQnA, TSpec-LLM, and SPEC5G (Nazar et al., 25 May 2025). ORAN-Bench-13K is described as approximately 13,000 entries drawn from 116 O-RAN specification documents, while TSpec-LLM covers 3GPP Releases 8–19 textual specifications (Nazar et al., 25 May 2025). Maintenance is incremental: new standards are periodically ingested through the same pipeline; FAISS indices are updated incrementally; and conflict detection plus automated consistency checks are applied for data quality (Nazar et al., 25 May 2025).

Evaluation in that paper uses answer relevancy, context recall, correctness, and faithfulness. The reported formulas are

$6.7 \times 10^9$4

$6.7 \times 10^9$5

$6.7 \times 10^9$6

and

$6.7 \times 10^9$7

Across five datasets with $6.7 \times 10^9$8 per set, the reported RAG-LLM results are high in recall and substantially improved over vanilla LLM baselines in correctness (Nazar et al., 25 May 2025). The paper states that LLaMa3.1-70B achieves a correctness score of $6.7 \times 10^9$9 and an answer relevancy rating of $P_{\mathrm{con}}$0 in the abstract, while the detailed quantitative table averaged across five datasets reports AR $P_{\mathrm{con}}$1, Recall $P_{\mathrm{con}}$2, AC $P_{\mathrm{con}}$3, and AF $P_{\mathrm{con}}$4 for LLaMa3.1-70B (Nazar et al., 25 May 2025). This discrepancy reflects two distinct reported summaries within the same source rather than a contradiction that can be resolved from the available text.

The paper also reports that “RAG versions dramatically outperform ‘vanilla’ LLMs (no retrieval) in correctness,” exemplified by vanilla Mixtral-8×7B at approximately $P_{\mathrm{con}}$5 versus RAG at approximately $P_{\mathrm{con}}$6, and that recall remains above $P_{\mathrm{con}}$7 across all RAG-LLMs (Nazar et al., 25 May 2025). This establishes knowledge grounding as one of the central technical identities of NextG-GPT.

4. Cloud–edge orchestration, personalization, and model adaptation

In the cloud–edge NetGPT architecture, personalization is introduced through prompt expansion at the edge. The edge LLM incorporates $P_{\mathrm{con}}$8, described as location- or user-specific context, to transform a concise prompt into a comprehensive prompt before deciding whether to answer locally or defer to the cloud (Chen et al., 2023). The edge can therefore act both as an inference node and as a contextualizer for cloud inference.

This workflow is tied to an AI-native Control/Compute/User Plane architecture. A Computing Plane sits alongside the traditional Control and User planes and dynamically assigns CPU, GPU, and VPU resources, schedules LoRA fine-tuning jobs, manages model-distribution and profile databases, and injects AI-prioritized flags into RRC or PDCP to reserve RAN resources for generative sessions (Chen et al., 2023). Feedback logs from edge and cloud inference are returned to the CmP for online profiling and future fine-tuning (Chen et al., 2023).

The same paper provides a stylized optimization problem for model placement and resource allocation. Let $P_{\mathrm{con}}$9 be active user sessions, $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$0 index edge or cloud, and $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$1 indicate where a session is served. The objective minimizes end-to-end latency under compute and bandwidth constraints, with compute-time and prompt-size terms appearing explicitly in the formulation (Chen et al., 2023). In practice, the paper notes that a softer convex relaxation can replace the binary assignment constraints and add penalties for model-switching overhead (Chen et al., 2023).

LoRA is the primary fine-tuning mechanism in this architecture. For a large weight matrix $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$2, LoRA freezes $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$3 and learns a low-rank update

$$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$4

where

$$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$5

The forward pass is written as

$$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$6

with $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$7 scaling the update (Chen et al., 2023). The reported resource effects are substantial: training only $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$8 and $$P_{\mathrm{com}} = \mathrm{LLM}_{\theta_e}(P_{\mathrm{con}}; I_{\mathrm{personal}})$$9 reduces VRAM from approximately $\rho$0 GB to approximately $\rho$1 GB for LLaMA-7B and shrinks stored fine-tuned parameters from $\rho$2 GB to approximately $\rho$3 MB at rank $\rho$4 (Chen et al., 2023).

The numerical comparison in that paper emphasizes latency and resource tradeoffs. For 100 prompts and 1 Gbps links, cloud-only latency is reported as $\rho$5 s, NetGPT synergy as $\rho$6 s, LLM splitting with $\rho$7 of layers at the edge as approximately $\rho$8 s, and full LLaMA offload at the edge as approximately $\rho$9 s but requiring $P_{\mathrm{com}}$0 GB storage and $P_{\mathrm{com}}$1 GB VRAM (Chen et al., 2023). NetGPT uses approximately $P_{\mathrm{com}}$2 GB VRAM at the edge for GPT-2-base prompt expansion and transfers approximately $P_{\mathrm{com}}$3 bytes per request on average, compared with approximately $P_{\mathrm{com}}$4 bytes for cloud-only (Chen et al., 2023). The paper also reports better personalization quality in location-aware tests such as tourism and library recommendations, and network-task performance including GPT-2-base popularity prediction accuracy of approximately $P_{\mathrm{com}}$5 versus LSTM at approximately $P_{\mathrm{com}}$6 and GRU at approximately $P_{\mathrm{com}}$7 (Chen et al., 2023).

5. Foundation-model formulation for wireless communications

The foundation-model perspective in “Ten issues of NetGPT” formalizes NetGPT as a pre-training and fine-tuning paradigm over wireless-domain token sequences (Tong et al., 2023). Data ingestion includes quantization or tokenization of continuous CSI into finite bins, sequence construction over control actions or channel observations, and supervised labeling such as optimal MCS labels or slice assignments (Tong et al., 2023). A cloud or HPC pre-training cluster hosts NetGPT-L0, supported by data-parallel and model-parallel engines and an objective module implementing autoregressive or masked modeling loss with domain regularizers (Tong et al., 2023).

The autoregressive pre-training objective is

$P_{\mathrm{com}}$8

where $P_{\mathrm{com}}$9 is implemented by a Transformer decoder and $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$0 penalizes large deviations from the random initialization priors $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$1 (Tong et al., 2023). An alternative masked-modeling objective is

$R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$2

with a domain regularizer

$R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$3

These objectives express the idea that NetGPT should learn not only language-like token dependencies but also structure specific to wireless channels, power control, or spectral characteristics (Tong et al., 2023).

For downstream network control, the paper introduces constrained fine-tuning over trajectories $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$4, maximizing expected spectral efficiency under power and latency constraints: $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$5 subject to

$R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$6

The corresponding Lagrangian is

$R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$7

Gradient updates are said to combine policy-gradient and constrained optimization steps (Tong et al., 2023).

The paper also frames design issues that remain central to the broader NextG-GPT discussion. “Data Collection & Labeling” highlights heterogeneity across vendors and cells; “Cross-Layer Collaboration” emphasizes standardized interfaces between edge and cloud models; and “Model Scalability & Distributed Deployment” addresses the inability to run very large NetGPT-L0 models on the edge without careful partitioning and federated updates (Tong et al., 2023). Recommended training protocols include bucketing by sequence length, mixed-precision input with 8-bit CSI tokens and 16-bit hidden activations, curriculum pre-training over mobility complexity, NCCL-based all-reduce for distributed training, tensor-slice model parallelism, and asynchronous federated updates for NetGPT-L2 (Tong et al., 2023).

This formulation broadens the scope of NextG-GPT beyond textual assistance. It positions the concept as a general-purpose foundation-model substrate for beamforming, network slicing orchestration, semantic communications, and link adaptation (Tong et al., 2023).

6. Agentic collaboration and reinforcement learning in NetGPT

The 2026 agentic framework is the most explicit attempt to make NetGPT a collaborative reasoning system for AI-native xG networks (Yu et al., 31 Jan 2026). It distinguishes an Autonomous Reasoning Module, used when tasks are low-complexity or when delegation latency is prohibited, from an Agent Invocation Module, which detects supported actions, decomposes intent into sub-tasks, queries the registry, routes to candidate agents, issues RPC or semantic messages, and integrates the returned answer into the core state (Yu et al., 31 Jan 2026).

Delegation is formalized through a communication protocol. The workflow proceeds through intent evaluation, decomposition into action invocation tokens such as <action type='NetworkAnalysis' goal='rootCause'> ... </action>, routing and selection over agent-card metrics, message dispatch using JSON or protoRPC with header: {task_id, action_type, context_embedding} and body: {goal_description, relevant_data}, agent execution returning <ans> ... informative result ... </ans>, and integration in which the core masks everything outside <ans> ... </ans> and incorporates the payload into its internal state (Yu et al., 31 Jan 2026). The system uses REST/gRPC for the agent registry and an ordered message queue on the NetGPT-Agent channel, cited as A2A/ACP/ANP standard (Yu et al., 31 Jan 2026).

Training is posed as a POMDP. Hidden states $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$8 contain all past tokens generated by NetGPT, the most recent agent responses with masked contents, and network-level observations such as latency and load (Yu et al., 31 Jan 2026). Actions $R = \mathrm{LLM}_{\theta_c}(P_{\mathrm{com}})$9 can be token emission over the vocabulary during autonomous reasoning or orchestrator actions of the form $C = 800$0 (Yu et al., 31 Jan 2026). Observations $C = 800$1 are partial views consisting of masked agent output, token logits, and summary network metrics; the transition kernel is stochastic because network conditions and agent responses are non-deterministic (Yu et al., 31 Jan 2026).

A masked maximum-likelihood loss is used during RL rollouts: $C = 800$2 where $C = 800$3 equals 1 when the token at time $C = 800$4 is part of NetGPT’s own policy decision and 0 otherwise (Yu et al., 31 Jan 2026). Entropy-guided exploration augments the trajectory objective: $C = 800$5 with high entropy triggering “branching” rollouts to sample diverse reasoning paths (Yu et al., 31 Jan 2026).

Rewards are explicitly multi-objective: $C = 800$6 The components are defined as

$C = 800$7

$C = 800$8

and

$C = 800$9

The paper argues that this formulation balances task quality, coordination efficiency, and resource usage, and notes that service providers can tune $C_0 = 80$0 to trade off responsiveness and resource cost (Yu et al., 31 Jan 2026).

The reported learning outcome is that SFT alone tends either to over-invoke agents or under-invoke them, while Agentic RL learns entropy-based thresholds: low entropy and simple intent lead to self-answering, whereas high entropy or complex domain leads to delegation (Yu et al., 31 Jan 2026). Empirically, the framework “invokes 1.8 agents on average per complex task (down from 3.4 before RL) while raising success rate from 62%→87%” (Yu et al., 31 Jan 2026). On TeleQnA and network root-cause tasks, the performance table reports the following: Pure Prompt with 45% success rate, 0.3 average agents invoked, 1200 ms average latency, and 1.0× compute cost; SFT Fine-Tuned with 62%, 1.5, 950 ms, and 1.2×; and Agentic RL (Ours) with 87%, 1.8, 780 ms, and 1.1× (Yu et al., 31 Jan 2026). Figure 1 is described as showing stable descent after 5k episodes, while ablations without entropy bonus or without masked loss fail to converge; Figure 2 reports internal reasoning up by 20%, agent calls down by 25% for mid-difficulty tasks, while preserving greater than 90% task success (Yu et al., 31 Jan 2026).

7. Applications, adjacent systems, and open questions

Across the literature, NextG-GPT and NetGPT are associated with several application classes. The foundation-model paper lists dynamic beamforming control, network slicing orchestration, semantic communications, and link adaptation (Tong et al., 2023). The RAG-focused NextG-GPT paper emphasizes research assistance, real-time support, and future adaptive RAN optimization, autonomous wireless experimentation, automated protocol validation, and security-vulnerability detection (Nazar et al., 25 May 2025). The cloud–edge paper adds personalized generative services, popularity prediction, intent inference, and intelligent network management and orchestration (Chen et al., 2023). The agentic NetGPT paper extends this to autonomous sensing, reasoning, and action in complex communication environments (Yu et al., 31 Jan 2026).

A related but distinct 2026 system, AgentxGCore, demonstrates how the broader agentic-AI trend is entering next-generation mobile core networks (Barbosa et al., 29 May 2026). It introduces an “Intelligent Layer” above 3GPP control-plane and user-plane functions, centered on an Intent Manager, a Network Planner Agent, a Network Executor Agent, and an MCP Server (Barbosa et al., 29 May 2026). Although not labeled NextG-GPT, it shares several motifs with agentic NetGPT: intent decomposition, specialized roles, tool registry mediation, telemetry-driven closed loops, and LLM-based policy execution (Barbosa et al., 29 May 2026). This suggests that the NextG-GPT trajectory is part of a wider movement toward AI-native network control stacks.

Several limitations recur across the sources. The RAG paper explicitly states that no formal hypothesis tests or confidence intervals were reported for its five-dataset results, only averages over 30 Q&A pairs per dataset (Nazar et al., 25 May 2025). The agentic NetGPT paper claims convergence behavior for the full method and failures for key ablations, but the summary does not provide a formal proof, and the training setting remains partially observable and stochastic by design (Yu et al., 31 Jan 2026). The foundation-model paper identifies unresolved issues in data heterogeneity, cross-layer collaboration, and distributed deployment (Tong et al., 2023). The cloud–edge paper highlights the need for AI-prioritized signaling, privacy protection, and workflow-aware orchestration at millisecond timescales (Chen et al., 2023). The related AgentxGCore work explicitly names security, data privacy, model availability, message overhead, cross-domain coordination, and lack of formal stability guarantees as open issues (Barbosa et al., 29 May 2026).

A common misconception is to treat NextG-GPT as only a telecom-domain chatbot. The literature is broader. In one strand, it is a RAG system for wireless-domain question answering (Nazar et al., 25 May 2025); in another, it is a collaborative cloud–edge architecture for personalized generative services (Chen et al., 2023); in another, it is a foundation-model program for wireless communications (Tong et al., 2023); and in the most recent strand, it is an agentic reasoning framework with POMDP-based reinforcement learning and explicit communication incentives (Yu et al., 31 Jan 2026). A second misconception is that the line of work is already standardized. The papers instead present architectural proposals, prototype workflows, training objectives, and experimental evidence rather than a single agreed protocol stack.

Taken together, the NextG-GPT literature describes an emerging class of AI-native network systems in which retrieval, model adaptation, edge–cloud coordination, and agentic delegation are increasingly unified. The central technical idea is consistent across its variants: network intelligence should be compositional, context-aware, resource-conscious, and integrated with the communication substrate itself (Nazar et al., 25 May 2025, Chen et al., 2023, Tong et al., 2023, Yu et al., 31 Jan 2026).