Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2511.22217

Optimizing NetGPT via Routing-Based Synergy and Reinforcement Learning

Yuxuan Chen·Rongpeng Li·Xianfu Chen·Celimuge Wu·Chenghui Peng·Zhifeng Zhao·Honggang Zhang
View original source ↗Read source document ↗Export metadata ↓

Abstract

Large language model (LLM) agents at the network edge offer low-latency execution for routine queries. In contrast, complex requests often require the superior capability of cloud models, incurring higher latency and cost. To navigate this quality-cost trade-off under dynamic network conditions, we propose a cloud-edge synergy for NetGPT that integrates network-aware routing with on-edge self-improvement. Specifically, our framework routes structured tool-calling requests to cloud or edge agents via a novel scoring policy. We prove that, under mild regularity assumptions, the optimal routing rule admits a unique fallback threshold with monotone dependence on bandwidth and round-trip time (RTT). Concurrently, based on the dataset collected from requests routed to the cloud and corresponding responses, we instantiate a schema-preserving reinforcement learning (RL) to improve the capability of the edge agent. We analyze a supervised finetuning (SFT)-anchored composite objective that combines a reverse-KL trust-region step with a forward-KL realignment toward the SFT prior, explaining stability and constraining policy drift. Both the network-aware routing policy and the edge agent are updated coherently. Experiments across controlled network states and pricing schedules demonstrate smooth quality-cost frontiers, consistent gains of dynamic fallback thresholds over fixed policies, and sustained reductions in offloading while maintaining task success and schema-correct outputs.

Explore related subjects

cs.NI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yuxuan Chen, Rongpeng Li, Xianfu Chen, Celimuge Wu, Chenghui Peng, Zhifeng Zhao, Honggang Zhang. 2025-11-27. Optimizing NetGPT via Routing-Based Synergy and Reinforcement Learning. https://arxiv.org/abs/2511.22217

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

CSI Simulation: Why Additive Noise Fails and How to Fix It

Channel State Information (CSI) has become a widely used wireless channel sensing modality for applications such as indoor localization, activity recognition, and respiration monitoring. Because collecting labeled data under every target condition is impractical, training CSI-based models often relies on simulated data produced by adding noise or perturbations to recorded channel estimates, most commonly additive white Gaussian noise (AWGN). This practice assumes that the receiver chain between the antenna and the channel estimator is linear and gain-invariant. We test this assumption empirically using RF jamming as a controlled perturbation on 6 commodity receivers across 2 indoor environments. The assumption does not hold. Automatic gain control compresses the channel estimate multiplicatively before digitization, producing amplitude distributions that no additive noise variance can reproduce. To close the resulting fidelity gap, we propose M_QTC, a measurement-calibrated model that learns the per-subcarrier distribution transformation through quantile mapping, temporal filtering, and copula-based cross-subcarrier reordering. M_QTC reduces amplitude error 8-fold and closes 89% of the aggregate fidelity gap across four complementary dimensions. The improvement transfers directly to downstream tasks, where 5 classifiers from different families trained on M_QTC-simulated data recover 93% of real-data jamming detection performance, while AWGN-trained classifiers remain near random decision.

cs.NI↗
arXiv2026

BALANCE: Hybrid Autoregressive-Speculative LLM Inference at the Network Edge

Edge inference is a promising paradigm to provide large language model (LLM) inference services in next-generation mobile networks. LLM inference mainly relies on two approaches: Autoregressive decoding (AD) generates output tokens sequentially, resulting in long latency; Speculative decoding (SD) accelerates inference by using a small language model (SLM) to generate multiple draft tokens for LLM verification, but incurs extra memory costs. Due to this latency-memory tradeoff, neither approach alone can efficiently serve users with heterogeneous demands under limited edge computing resources. To address this challenge, we propose a hybrid autoregressive-speculative inference (BALANCE) framework for edge LLM inference. In BALANCE, an edge server hosts both an SLM and an LLM, admits users, assigns each admitted user to the AD or SD mode, and performs the two modes simultaneously. To maximize the number of served users, we formulate a task throughput maximization problem to jointly determine user admission and computing resource allocation between AD and SD under user latency requirements and server memory constraints. Since the problem is NP-hard, we develop a polynomial-time algorithm that transforms the original problem into two sub-problems and obtains a sub-optimal solution with a constant approximation guarantee. Experiments demonstrate that BALANCE consistently outperforms conventional AD and SD and significantly improves task throughput.

cs.NI↗
arXiv2026

Zero-Knowledge Remote Adversarial Attack against Wi-Fi-based Human Activity Recognition for Privacy Protection

The growing capability of Wi-Fi devices to identify human activities using channel state information (CSI) raises privacy concerns. To counter this threat, we propose GRAW, an adversary system, acting as a privacy defender, that degrades the human activity recognition (HAR) system at the user device by perturbing the router's signals that the device uses to estimate CSI. GRAW employs generative adversarial imitation learning (GAIL) to construct perturbation signals, and thereby eliminates the need for any information on the target HAR systems and their inputs (i.e., zero-knowledge operation). We evaluate GRAW against seven representative HAR models, using datasets collected in five environments, including our own dataset. We observe that GRAW is the only remote attack scheme that degrades every tested HAR model to a random-selection level. At the same perturbation level, GRAW achieves an attack success ratio up to 76.7% higher than comparison methods, while maintaining over 99% packet success rate on regular Wi-Fi communication. We demonstrate the feasibility of GRAW through real-time, over-the-air experiments with software-defined radios.

cs.NI↗