AIR-LLM is an edge inference architecture that broadcasts large language model (LLM) weights over radio, allowing edge devices to perform matrix-vector multiplications directly in the RF domain without storing or loading the weights. The system uses MIMO spatial multiplexing and an energy‑efficient precoder‑postcoder pair to reduce airtime and calibrate the wireless channel, enabling a single broadcast to serve unlimited users. Experiments on real urban channel models show that AIR-LLM achieves only a 4.0% perplexity loss on LLaMA‑3.1‑8B while saving energy by up to 157.7× compared to FP16 and reducing airtime by over 100× for 20 users.
By Zhihui Gao, Tingjun Chen, Dirk Englund
arXiv:2609.08312v1 Announce Type: cross
Abstract: To mitigate the scalability bottleneck in the radio access network (RAN) in federated edge learning (FEEL), over-the-air federated learning (AirFL) e...
By Haifeng Wen, Nicol\`o Michelusi, Osvaldo Simeone, Yang Yang, Hong Xing
The paper proposes using analog radio frequency (RF) computing to enable edge devices to perform AI inference by receiving a waveform that encodes neural network weights from a base station. Clients then mix this waveform with their own input-encoded signals to perform matrix‑vector multiplications with minimal power and memory. In a multi‑user MIMO downlink setting, the authors formulate a joint beamforming optimization to balance data and AI flows, showing that the aggregate data rate loss is below 5% under realistic 3GPP channel conditions.
By Wentao Yu, Vincent W. S. Wong
arXiv:2509. 16577v2 Announce Type: replace Abstract: Over-the-air (OTA) aggregation enables federated edge learning (FEEL) by exploiting the superposition property of the wireless channel to merge communication with computation, eliminating the need to schedule and decode devices individually.
By Antonio Tarizzo, Mohammad Kazemi, Deniz G\"und\"uz
The paper proposes a machine‑learning–based wireless image transmission scheme that merges compression and channel coding via a multi‑view autoencoder, enabling multiple transmitters to share the same channel resources simultaneously in a non‑orthogonal multiple‑access (NOMA) setting. A progressive fine‑tuning algorithm doubles the number of users at each iteration while preserving orthogonalized performance, allowing the system to support up to 64 users with only a 2.4 % increase in trainable parameters. The method improves recovered image quality and outperforms existing NOMA‑based approaches across various datasets, metrics, and channel conditions.
By Selim F. Yilmaz, Can Karamanli, Deniz Gunduz
arXiv:2602. 11834v2 Announce Type: replace-cross Abstract: While machine learning (ML)-based receiver algorithms have received a great deal of attention in the recent literature, they often suffer from poor scaling with increasing spatial multiplexing order and lack of explainability and generalization.
By Mikko Honkala, Dani Korpi, Elias Raninen, Janne M. J. Huttunen