Research
Latest academic paper digests, data charts, and research insights from the 6G and AI research community.
16 papers
State Space Models for Wireless Channel Prediction
Dr. Lukas Mauch, Prof. Hans Schotten — RPTU Kaiserslautern / 6G Research Hub
We apply structured state space models (S4/Mamba) to wireless channel prediction, demonstrating that these architectures offer superior long-range sequence modeling compared to transformers and LSTMs for time-varying channels. Our S4-Channel model predicts channel state information 10-50ms into the future with 40% lower prediction error than transformer baselines, while requiring linear rather than quadratic computation in sequence length. This enables predictive beamforming and proactive resource allocation.
Distributed Foundation Models for Heterogeneous Network Optimization
Dr. Mingzhe Chen, Prof. Walid Saad et al. — Virginia Tech / Beijing University of Posts and Telecommunications
We introduce a distributed foundation model framework for optimizing heterogeneous wireless networks comprising macro cells, small cells, and Wi-Fi access points. A pre-trained base model is split across network tiers with tier-specific adapter modules, enabling coordinated optimization without sharing raw data between tiers. The framework achieves 30% better network-wide throughput than independent per-tier optimization while reducing inter-tier interference by 45%.
Physics-Informed Neural Networks for Radio Propagation Modeling
Dr. Andreas Molisch, Dr. Dawei Ying — University of Southern California
We develop physics-informed neural networks (PINNs) for radio propagation modeling that incorporate Maxwell's equations as soft constraints during training. By encoding electromagnetic wave physics directly into the loss function, our PINNs predict path loss and multipath characteristics with 5x less training data than purely data-driven approaches while maintaining equivalent accuracy. The model generalizes to unseen environments 3x better than standard neural network propagation models.
Token-Free Language Models for Efficient Telecom Log Analysis
Dr. Chen Li, Dr. Marco Fiore — IMDEA Networks / NEC Laboratories Europe
Traditional LLMs struggle with telecom network logs due to their technical vocabulary and structured format not aligning well with standard tokenization. We propose a byte-level token-free language model specifically designed for telecom log analysis. Our model processes raw byte sequences directly, avoiding out-of-vocabulary issues common with standard tokenizers on network log data. On a benchmark of 1M real operator logs, our approach achieves 91% fault classification accuracy and generates root cause explanations that experts rate as helpful 85% of the time.
Reward Shaping for Safe Reinforcement Learning in Network Control
Dr. Tianyu Wang, Prof. Robert Schober — University of Erlangen-Nuremberg
Deploying RL agents in live networks carries the risk of unsafe actions that degrade service. We propose a reward shaping framework that incorporates safety constraints from network SLAs directly into the RL training process. Our constrained RL approach guarantees that QoS violations remain below 0.1% while still achieving 90% of the throughput optimality of unconstrained agents. We validate on a commercial 5G testbed with 50 active users.
Vision Transformers for RF Spectrum Monitoring and Classification
Dr. Tim O'Shea, Dr. Nathan West — DeepSig Inc.
We apply Vision Transformers (ViT) to RF spectrum monitoring by treating spectrograms as images. Our ViT-RF model classifies 24 modulation types with 98.5% accuracy at 10 dB SNR, outperforming CNN baselines by 3.2%. The attention mechanism provides interpretable visualization of which time-frequency regions drive classification decisions. The model runs at 2ms per spectrogram on edge GPU hardware, enabling real-time spectrum monitoring for 6G cognitive radio applications.
Efficient On-Device Training for Adaptive 6G Receivers
Dr. Junmo Kim, Prof. Youngchul Sung — KAIST
We propose an efficient on-device training framework that enables 6G receiver neural networks to continuously adapt to changing channel conditions without cloud connectivity. Using a combination of pruned backpropagation and mixed-precision training, our approach reduces on-device training memory by 8x and energy by 5x compared to standard backpropagation, while maintaining adaptation quality. This enables continuous learning on mobile device chipsets with limited resources.