NetGPT: Generative AI in Networked Systems

• NetGPT is a family of generative pretrained transformers adapted for network-centric applications across wireless communications, traffic modeling, and graph-based analysis.
• It employs techniques such as low-rank adaptation and agentic reinforcement learning to optimize resource use and reduce latency in cloud-edge orchestration.
• Empirical evaluations demonstrate improved task accuracy, coordination efficiency, and scalability across diverse network environments in real-world deployments.

NetGPT is a designation used in several arXiv research lines at the intersection of generative pretrained models and networked systems. In wireless communications, it denotes AI-native network architectures that combine edge and cloud LLMs, a broader family of foundation models for wireless systems, and later agentic frameworks for collaborative reasoning in xG networks (Chen et al., 2023, Tong et al., 2023, Yu et al., 31 Jan 2026). The same designation has also been used for a generative pretrained transformer for network traffic understanding and generation (Meng et al., 2023), and for a large graph model for short-video propagation influence rating (Xue et al., 31 Mar 2025). Across these uses, the common emphasis is on coupling pretrained generative models with structured network state, resource constraints, or graph-structured observations.

1. Terminological scope and conceptual lineage

In the wireless-communications literature, NetGPT is explicitly defined as “Network Generative Pre-trained Transformer,” namely a class of foundation models for wireless communications rather than a single monolithic model. “Ten issues of NetGPT” states that NetGPT is a family of transformers of various sizes and specializations tailored to domains such as RAN, CN, OAM, and UE, and organizes this family into a three-layer hierarchy: NetGPT-L0, NetGPT-L1, and NetGPT-L2 (Tong et al., 2023). This directly counters a common misconception that NetGPT refers to one fixed architecture.

The surveyed literature also uses the same name for distinct technical artifacts. In 2023, one line presented NetGPT as a native-AI network architecture for cloud-edge personalized generative services (Chen et al., 2023). A later xG-oriented line proposed an agentic NetGPT core with domain-specialized agents and agentic reinforcement learning (Yu et al., 31 Jan 2026). Outside wireless control and orchestration, another work used NetGPT for a GPT-based pretrained model over network traffic byte streams (Meng et al., 2023), while a separate line used NetGPT for a large graph model operating on a cross-platform short-video propagation graph (Xue et al., 31 Mar 2025).

Usage of “NetGPT”	Domain	Representative work
Network foundation-model family	Wireless communications	(Tong et al., 2023)
Cloud-edge personalized generative architecture	AI-native network architecture	(Chen et al., 2023)
Agentic collaborative reasoning framework	xG wireless networks	(Yu et al., 31 Jan 2026)
Generative pretrained transformer for traffic	Network traffic modeling	(Meng et al., 2023)
Large graph model	Short-video propagation analysis	(Xue et al., 31 Mar 2025)

A plausible implication is that “NetGPT” functions more as a research motif than as a single stable specification: the term consistently signals the transplantation of GPT-like or LLM-like mechanisms into network-centric problem domains, but the underlying objects differ substantially.

2. Cloud-edge NetGPT for personalized generative services

“NetGPT: A Native-AI Network Architecture Beyond Provisioning Personalized Generative Services” proposes a collaborative cloud-edge methodology in which an edge LLM and a cloud LLM are orchestrated according to computing capacity and personalization requirements (Chen et al., 2023). The edge tier is deployed on a base-station or edge server and uses GPT-2-base (approximately $0.1$ B parameters). Its responsibilities are prompt enhancement or completion through location and personalized information, preliminary de-duplication and buffering of requests, and local inference for trend prediction and intent inference as by-products. The cloud tier is deployed in centralized data centers and uses a fine-tuned LLaMA-7B model (approximately $6.7$ B parameters) for final generative response and heavy-weight inference and RLHF loops. An orchestration module maintains a resource database, latency/performance profiles of each LLM, and user profiles with location metadata, and performs resource allocation, workflow mapping, and offloading decisions.

The service pipeline begins with a concise prompt $P_{\mathrm{con}}$ from the user. At the edge, the model computes a comprehensive prompt

$$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$

where $I_{\mathrm{personalized}}$ includes base-station location from the 5G AMF and venue or facility metadata. The paper reports $P_{\mathrm{com}}$ at roughly $95$ bytes versus $12$ bytes for the concise prompt (Chen et al., 2023). The logical workflow comprises request arrival, buffering and de-duplication, edge inference, optional early return if local confidence exceeds $\tau_{\mathrm{local}}$, cloud inference otherwise, post-processing, delivery, a feedback loop, and continuous orchestration updates based on latency, accuracy, and throughput.

The same work also formalizes cloud-edge resource orchestration as a binary decision problem over requests $x_i \in \{0,1\}$, minimizing expected latency plus resource cost subject to edge and cloud CPU/GPU capacity and throughput constraints (Chen et al., 2023). This formulation anchors NetGPT not merely as an LLM deployment pattern but as an architecture-level proposal for integrating communication and computing resources in AI-native networks.

3. Agentic NetGPT and collaborative reasoning in xG networks

“Communications-Incentivized Collaborative Reasoning in NetGPT through Agentic Reinforcement Learning” re-specifies NetGPT as a unified agentic framework for AI-native xG networks (Yu et al., 31 Jan 2026). Its NetGPT Core is built on a standard LLM backbone such as LLaMA-3 Instruct and extends it with an Intent Interpreter, a Reasoning Module, and a Coordination Controller. The Core maintains a protocol-agnostic Agent Registry containing agent cards with supported actions, endpoint metadata such as latency, cost, and load, and identity and capability descriptors mapped from A2A, ACP, and ANP via adapters. Domain-specialized agents live in the user, control, or compute plane, or on an edge node, and comprise a Perception Interface, a Local Reasoner or Toolbox, and a Communication Adapter. A shared infrastructure layer exposes knowledge bases, data lakes, registry services, and persistent memories to both Core and agents.

The communication protocol is explicitly agentic. Supported protocols are A2A, ACP, and ANP via a thin adapter. Invocation messages use an <action ...> envelope, while agent replies use an <ans ...> envelope. A central masking rule states that only tokens between <ans> and </ans> are unmasked and fed back into the Core’s reasoning loop; all other tokens are masked out for policy-gradient updates (Yu et al., 31 Jan 2026). Under the paper’s Markov decision process, the state is

$6.7$0

where $6.7$1 is the decoder hidden state and $6.7$2 is the partial external observation returned by agents. The action space includes ordinary token emission and invocation actions:

$6.7$3

The collaboration objective is explicitly multi-objective:

$6.7$4

with $6.7$5, $6.7$6, and $6.7$7 (Yu et al., 31 Jan 2026). The decision mechanism compares the maximum internal-token probability with the maximum invocation probability, selecting an invocation when $6.7$8, where $6.7$9 is learned by an auxiliary head on a held-out validation set of “should-invoke” versus “should-reason-internally” labels. This formulation makes collaboration endogenous to the decoder policy rather than a purely external scheduler.

This suggests a broadening of NetGPT from earlier cloud-edge synergy toward explicit multi-agent delegation and partially observable coordination, although the surveyed papers present these as separate formulations rather than as a single standardized lineage.

4. Training and optimization paradigms

A recurring technical theme is parameter-efficient adaptation. In the cloud-edge architecture, low-rank adaptation is used for fine-tuning open-source LLMs. For a pretrained weight matrix $P_{\mathrm{con}}$0, the update is

$P_{\mathrm{con}}$1

with $P_{\mathrm{con}}$2, $P_{\mathrm{con}}$3, and $P_{\mathrm{con}}$4 (Chen et al., 2023). For LLaMA-7B, the paper reports VRAM reduction from $P_{\mathrm{con}}$5 GB to approximately $P_{\mathrm{con}}$6 GB and storage of fine-tuned parameters from approximately $P_{\mathrm{con}}$7 GB to approximately $P_{\mathrm{con}}$8 MB for $P_{\mathrm{con}}$9 and $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$0.

The agentic xG framework adopts a two-phase end-to-end loop consisting of supervised fine-tuning followed by agentic reinforcement learning (Yu et al., 31 Jan 2026). Phase 1 minimizes

$$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$1

over direct Q-A samples and action-driven dialogues with annotated <action>...</action> spans. Phase 2 performs rollouts in a simulated or real network environment, routes invoke(a) steps to concrete agents, computes multi-objective rewards, adds entropy-guided exploration, and applies a masked policy gradient. The paper gives $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$2, $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$3, $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$4, reward weights $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$5, $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$6, $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$7, and an annealing schedule for $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$8 down to $$P_{\mathrm{com}} = LLM^{\mathrm{edge}}_{\theta}(P_{\mathrm{con}}; I_{\mathrm{personalized}})$$9 (Yu et al., 31 Jan 2026).

A separate optimization line, “Optimizing NetGPT via Routing-Based Synergy and Reinforcement Learning,” formalizes cloud-edge routing as a threshold policy over a router score $I_{\mathrm{personalized}}$0 and a fallback threshold $I_{\mathrm{personalized}}$1 that depends on network state $I_{\mathrm{personalized}}$2 (Chen et al., 27 Nov 2025). Under assumptions that $I_{\mathrm{personalized}}$3 and that the benefit-cost ratio $I_{\mathrm{personalized}}$4 is continuous and strictly decreasing, the utility $I_{\mathrm{personalized}}$5 is maximized at the unique threshold $I_{\mathrm{personalized}}$6 satisfying

$I_{\mathrm{personalized}}$7

The paper derives monotone comparative statics: $I_{\mathrm{personalized}}$8 increases with bandwidth and decreases with RTT. It then gives two deployable parameterizations, FuncDyn and PolicyNet, and couples them with an SFT-anchored RL scheme whose composite objective combines a reverse-KL trust region against the old policy with a forward-KL realignment toward the SFT prior (Chen et al., 27 Nov 2025). In that formulation, the forward-KL term is explicitly intended to preserve schema correctness and prevent catastrophic forgetting of the tool-calling format.

5. Empirical evaluation in AI-native networking

The empirical literature on wireless NetGPT evaluates latency, coordination, cost, quality, and personalization under different formulations. The 2023 cloud-edge paper reports a numerical comparison for $I_{\mathrm{personalized}}$9 concise prompts at $P_{\mathrm{com}}$0 Gbps and batch $P_{\mathrm{com}}$1 (Chen et al., 2023). In that table, cloud-only yields end-to-end latency $P_{\mathrm{com}}$2 s with $P_{\mathrm{com}}$3 GB cloud VRAM and $P_{\mathrm{com}}$4 bytes transferred per request; LLM Offload yields $P_{\mathrm{com}}$5 s latency with $P_{\mathrm{com}}$6 GB edge VRAM; LLM Splitting ($P_{\mathrm{com}}$7) yields $P_{\mathrm{com}}$8 s latency with approximately $P_{\mathrm{com}}$9 bytes transferred; and NetGPT (Synergy) yields $95$0 s latency, $95$1 GB edge VRAM, $95$2 GB edge storage, and $95$3 bytes transferred per request. The same paper interprets these numbers as reducing edge VRAM and storage by more than $95$4 versus LLM Offload and reducing latency by approximately $95$5 versus cloud-only, while Fig. D reports $95$6–$95$7 higher BLEU/ROUGE for personalized tasks than cloud-only (Chen et al., 2023).

The 2026 agentic paper evaluates task accuracy $95$8, coordination efficiency $95$9, latency or SLA violation rate, and a composite score $12$0 on TeleQnA and TeleLogs (Yu et al., 31 Jan 2026). Its reported “numbers illustrative” are: Pure Prompting with $12$1, $12$2, Composite $12$3; SFT Only with $12$4, $12$5, Composite $12$6; and NetGPT + Agentic RL with $12$7, $12$8, Composite $12$9. The paper summarizes this as an $\tau_{\mathrm{local}}$0 absolute gain in accuracy over SFT, $\tau_{\mathrm{local}}$1 fewer unnecessary agent calls versus SFT-only, and a $\tau_{\mathrm{local}}$2 reduction in average end-to-end latency (Yu et al., 31 Jan 2026).

The 2025 routing-and-RL paper evaluates GOOD, MID, and BAD link regimes with RTT and bandwidth sampled from $\tau_{\mathrm{local}}$3 ms and $\tau_{\mathrm{local}}$4 Mbps, $\tau_{\mathrm{local}}$5 ms and $\tau_{\mathrm{local}}$6 Mbps, and $\tau_{\mathrm{local}}$7 ms and $\tau_{\mathrm{local}}$8 Mbps, respectively, under a Gauss-Markov model (Chen et al., 27 Nov 2025). Its reported findings are qualitative but precise: dynamic controllers such as FuncDyn and PolicyNet dominate fixed-threshold and learned static routers on the quality-cost frontier; dynamic thresholds adapt smoothly under regime switches; empirical frontiers $\tau_{\mathrm{local}}$9 versus $x_i \in \{0,1\}$0 are smooth and diminishing-returns; scanning $x_i \in \{0,1\}$1 shows a unique unimodal peak $x_i \in \{0,1\}$2 that falls as the network degrades or as cost-sensitivity $x_i \in \{0,1\}$3 increases; and PPO with SFT anchoring improves reward and reduces offload rate without spuriously violating JSON schema, whereas PPO without SFT anchoring is unstable (Chen et al., 27 Nov 2025).

These results are not directly commensurate because the tasks, metrics, and experimental environments differ. What they jointly establish is that NetGPT research in networking has moved from static cloud-edge partitioning toward adaptive routing, policy learning, and collaboration-aware utility optimization.

6. Non-wireless uses of the name: traffic pretraining and large-graph reasoning

One distinct usage appears in “NetGPT: Generative Pretrained Transformer for Network Traffic,” which presents a GPT-2-Base model for traffic understanding and generation (Meng et al., 2023). The backbone has $x_i \in \{0,1\}$4 Transformer layers, hidden size $x_i \in \{0,1\}$5, $x_i \in \{0,1\}$6 causal self-attention heads, and feed-forward inner dimension $x_i \in \{0,1\}$7. Input representation is a uniform multi-pattern encoding in which each raw byte is converted to a two-digit ASCII hex string and tokenized with WordPiece; the model uses special tokens [cls] and [pck], adds segment embeddings for flow-level fine-tuning, and remains strictly autoregressive. Fine-tuning employs header-field shuffling, packet segmentation across up to the first three packets of a flow, and prompt-based task labels that recast classification and generation as conditional generation. On five datasets totaling approximately $x_i \in \{0,1\}$8 GB raw PCAP, the paper reports packet-level average Accuracy/Macro-F1 of $x_i \in \{0,1\}$9 and flow-level average $6.7$00, outperforming ET-BERT and GPT-2 baselines; for traffic generation, the average Jensen-Shannon divergence across four test datasets is $6.7$01 for NetGPT versus $6.7$02 for GPT-2, with packet-length JSD on ISXW falling from $6.7$03 to $6.7$04 (Meng et al., 2023).

Another distinct usage appears in “Short-video Propagation Influence Rating: A New Real-world Dataset and A New Large Graph Model,” where NetGPT denotes a large graph model for the SPIR task (Xue et al., 31 Mar 2025). Here the objective is to predict a short video’s long-term influence level $6.7$05 within a cross-platform propagation graph. The XS-Video dataset comprises $6.7$06 videos, $6.7$07 samples, $6.7$08 topics, and approximately $6.7$09 nodes with approximately $6.7$10 edges. NetGPT combines a heterogeneous GNN encoder, an off-the-shelf LLM such as Qwen2-VL, and a three-stage training pipeline of Heterogeneous Graph Pretraining, Supervised Language Fine-tuning, and Task-oriented Predictor Fine-tuning. On the XS-Video test set, the reported performance is ACC $6.7$11, MSE $6.7$12, and MAE $6.7$13, compared with RGCN at $6.7$14 and Qwen2-VL at $6.7$15 (Xue et al., 31 Mar 2025).

These non-wireless uses matter for terminology. A common misconception is that NetGPT names only a wireless-network control architecture. The literature surveyed here shows that it also names a byte-level traffic pretrained model and a graph-LLM hybrid for social propagation analysis.

7. Open issues, limitations, and research questions

“Ten issues of NetGPT” frames the principal unresolved questions for the wireless line of research (Tong et al., 2023). It argues that dedicated AI models in wireless suffer from low generality, limited performance gain, complicated management and collaboration, and lack of multi-task processing capability. Against that background, the paper identifies ten issues spanning scenarios and requirements, theoretical gaps between NetGPT and LLMs, ultimate performance, collaboration mechanisms, distributed deployment, network architecture design, security and privacy, data governance, evaluation and SLA metrics, and full-lifecycle management and orchestration.

Several of these issues expose persistent technical tensions. The paper notes that communication data are continuous, high-dimensional tensors rather than ordinary text tokens, while outputs may be bit-level decisions, numeric predictions, or control commands rather than natural-language continuations (Tong et al., 2023). It also highlights stringent 6G objectives, including approximately $6.7$16 ms real-time inference and ultra-reliability, and states that hallucination cannot be tolerated in mission-critical network decisions. On the systems side, it points to heterogeneous end-edge-cloud hardware, intermittent connectivity, the need for dynamic partitioning, and the absence of standardized model update interfaces. On the governance side, it identifies backdoor and poisoning risks, exposure of sensitive user data during inference, the need for differential privacy or encryption in model updates, and the requirement for unified data-governance frameworks and cross-vendor evaluation suites.

These concerns should not be read as rejections of NetGPT. Rather, they delimit the conditions under which NetGPT-style systems would need to operate if they are to move from conceptual architecture and controlled evaluation toward telecom-grade deployment. The open questions are therefore not peripheral; they define the boundary between current demonstrations and operational AI-native networks.