This collection of preprints explores cutting-edge advancements in signal processing, communication systems, and sensing technologies, with a strong emphasis on applying machine learning and optimization techniques. Several papers delve into the challenges and opportunities presented by extremely large-scale antenna arrays (ELAAs) and millimeter-wave/terahertz communications. Galappaththige et al. (2025) propose a low-complexity algorithm based on Riemannian stochastic gradient descent and augmented Lagrangian manifold optimization (SGALM) for multi-target detection and communication in ELAA-based ISAC systems. Their SGALM approach achieves significant runtime improvements, addressing the practical limitations of conventional algorithms. Meanwhile, Tian et al. (2025) introduce a hybrid dynamic subarray (HDS) architecture and associated algorithms (RD-MUSIC) for efficient direction-of-arrival (DOA) estimation in THz UM-MIMO systems, tackling the complexities of large-scale antenna arrays and high-frequency operation. Complementing these ELAA-focused works, Liu et al. (2025) propose Pinching Antenna Systems (PASS), a novel architecture employing dynamic pinching beamforming for efficient short-range communication, demonstrating potential performance gains over conventional multi-antenna systems.
The integration of sensing and communication is another prominent theme. Akçalı et al. (2025) investigate predictive beamforming in distributed MIMO for V2X applications, leveraging sensing information to improve beam tracking and demonstrating the advantages of distributed antennas for robust sensing and communication performance. Han et al. (2025) tackle the challenges of noncoherent distributed ISAC, proposing a signal design framework for joint communication and MIMO radar operation without requiring phase synchronization, optimizing target localization while meeting communication SINR requirements. Rajwan & Boag (2025) develop a fast backprojection algorithm for near-field SAR imaging of moving targets, enhancing efficiency in detecting both moving and stationary vehicles. Müller et al. (2025) investigate low-complexity event detection and identification in coherent correlation OTDR measurements, demonstrating accurate localization and extraction of vibration parameters.
Machine learning plays a crucial role across several contributions. Choi & Park (2025) introduce VoicePrompter, a zero-shot voice conversion model leveraging in-context learning with voice prompts and conditional flow matching, achieving improved speaker similarity and naturalness. Chen et al. (2025) propose RadioLLM, integrating LLMs into cognitive radio through hybrid prompt and token reprogramming, demonstrating improved performance in tasks like radio signal classification. Sharma et al. (2025) utilize variational autoencoders (VAEs) for drivetrain simulation, generating realistic jerk signals from torque demands. Li et al. (2025) present an efficient stochastic polar decoder with correlated stochastic computing.
Beyond these core themes, the preprints explore diverse topics such as private information retrieval (Meel et al., 2025), over-the-air computation (Sato & Ishibashi, 2025), 5G channel modeling for railways (Guan et al., 2025), UAV air-ground channel modeling (Cui et al., 2025), and physical layer authentication (Crosara et al., 2025). Finally, several papers address practical implementation and system-level considerations, including machine learning fairness for depression detection (Kwok et al., 2025), personalized federated learning (Cooper & Geller, 2025), battery health estimation (Yunusoglu et al., 2025), power-efficient over-the-air aggregation (Moradi Kalarde et al., 2025), and explainable beam alignment (Khan et al., 2025).
RadioLLM: Introducing Large Language Model into Cognitive Radio via Hybrid Prompt and Token Reprogrammings by Shuai Chen, Yong Zu, Zhixi Feng, Shuyuan Yang, Mengchang Li, Yue Ma, Jun Liu, Qiukai Pan, Xinlei Zhang, Changjun Sun https://arxiv.org/abs/2501.17888
Caption: The RadioLLM framework processes a noised radio signal through instance normalization, Hybrid Prompt and Token Reprogramming (HPTR), and Frequency Attuned Fusion (FAF) before feeding it to a Large Language Model (LLM). The LLM, augmented by HPTR and FAF, classifies the signal and reconstructs a denoised version, demonstrating a novel approach to cognitive radio tasks. The framework components, including the prompt structure, token reprogramming, frequency feature fusion, and LLM architecture, are detailed in the diagram.
RadioLLM presents a paradigm shift in Cognitive Radio Technology (CRT) by introducing a unified framework capable of handling diverse tasks and signal types. The ever-increasing demand for wireless connectivity coupled with the finite nature of spectrum resources necessitates innovative solutions for efficient radio network management. Traditional deep learning approaches, while promising, often lack the adaptability and scalability required for the dynamic real-world radio frequency landscape. This paper addresses this challenge by leveraging the remarkable generalization capabilities of Large Language Models (LLMs), offering a versatile and scalable solution for a range of radio signal processing challenges.
Two key innovations drive the RadioLLM framework: Hybrid Prompt and Token Reprogramming (HPTR) and Frequency Attuned Fusion (FAF). HPTR bridges the gap between the abstract world of LLMs and the intricacies of radio signals. It achieves this by combining expert-crafted prompts with semantically similar anchors derived from signal embeddings. This approach effectively injects domain-specific knowledge into the LLM while minimizing computational overhead. FAF, on the other hand, addresses a critical limitation of LLMs: their difficulty in capturing high-frequency signal features crucial for precise signal processing. FAF accomplishes this by fusing high-frequency features extracted by Convolutional Neural Networks (CNNs) with the LLM's broader contextual understanding. A lightweight decoder then maps these processed features back to the original signal space, enabling tasks like denoising and signal recovery. The received signal, r(t), can be modeled as:
r(t) = h(t) * s(t) + n(t)
where s(t) represents the transmitted signal, h(t) is the channel response, and n(t) is additive white Gaussian noise.
Rigorous empirical studies on benchmark datasets like RML2016a, RML2016b, RML2016c, RML2018a, ADS-B, and Wi-Fi demonstrate RadioLLM's superior performance. In radio signal classification (RSC), it consistently outperforms state-of-the-art methods, showcasing its adaptability and generalization prowess. For example, on RML2016a, RadioLLM achieves higher overall accuracy than competing methods in both 50-shot and 100-shot learning settings. Its effectiveness extends to denoising tasks as well, achieving the highest Structural Similarity Index Measure (SSIM) across all datasets, preserving signal integrity while effectively removing noise. Ablation studies further confirm the significant contributions of both HPTR and FAF modules. RadioLLM sets a new precedent for intelligent and adaptable wireless communication networks, promising more efficient management of the increasingly complex radio frequency spectrum.
Advancing Brainwave-Based Biometrics: A Large-Scale, Multi-Session Evaluation by Matin Fallahi, Patricia Arias-Cabarcos, Thorsten Strufe https://arxiv.org/abs/2501.17866
Caption: This graph displays the Equal Error Rate (EER) for brainwave authentication over various time intervals, ranging from one day to three years, using the PEERS dataset. The number of subjects in each time interval is also shown, illustrating the varying sample sizes used for evaluation. The results highlight the challenge of performance degradation over time, with EER increasing significantly from one day to one year.
Brainwave-based biometrics offers a compelling alternative to traditional authentication methods, providing advantages such as hands-free interaction, resistance to shoulder surfing, and the potential for continuous authentication. This study marks a significant advancement in the field by addressing a key limitation of prior research: the reliance on small, single-session datasets. Leveraging the massive, publicly available PEERS dataset – encompassing 345 subjects and over 6,000 sessions collected over five years with three different EEG headsets – this research provides a far more robust and realistic evaluation of brainwave biometrics in a multi-session context.
The study conducted a comprehensive benchmark of various feature extraction and comparison methods. Supervised Contrastive Learning (SupConLoss), a deep learning approach, coupled with a simple cosine distance metric, emerged as the top-performing pipeline, achieving an Equal Error Rate (EER) of 10.75%. This significantly outperforms classic feature extraction methods like Power Spectral Density (PSD), which resulted in an EER of 27.3%. Importantly, the study found that increasing the number of test subjects did not significantly impact EER, indicating the robustness of this metric to larger evaluation sets.
A crucial finding of the study is the degradation of authentication accuracy over time. The EER increased from 7.7% after one day to 19.69% after one year, underscoring the necessity for continuous enrollment updates after successful logins to maintain system integrity. Visualization of the embedding space revealed a clear session effect, with samples from the same session clustering together. This suggests that factors like environmental noise, device setup, and physiological changes in the subject over time contribute to intra-subject variability. Cross-hardware authentication was also explored, highlighting the importance of training the feature extractor with cross-hardware sessions for acceptable performance. Furthermore, the study investigated the impact of reducing the number of EEG channels, simulating the transition to more affordable consumer-grade devices. The results showed a manageable increase in EER – from 10.78% with 93 channels to 13.97% with 14 channels – suggesting that brainwave authentication is feasible with less complex hardware.
While the achieved performance surpasses previous work and demonstrates promise, it still falls short of stringent industry standards set by NIST and ISO. However, the researchers observed a linear relationship between the error rate and the logarithm of the number of training subjects. This trend suggests that reaching industry standards is feasible with a larger training dataset of at least 1,500 subjects. The open-sourcing of the analysis code further promotes continued research and development in this exciting field.
Drivetrain simulation using variational autoencoders by Pallavi Sharma, Jorge-Humberto Urrea-Quintero, Bogdan Bogdan, Adrian-Dumitru Ciotec, Laura Vasilie, Henning Wessels, Matteo Skull https://arxiv.org/abs/2501.17653
Caption: The figure compares original and generated jerk signals (a), their corresponding spectrograms (b, c), and the absolute error between them (d). The generated signal and spectrogram are produced by a variational autoencoder (VAE) trained on experimental data from electric SUVs, demonstrating the VAE's ability to replicate real-world drivetrain behavior.
This research explores a novel approach to drivetrain simulation using variational autoencoders (VAEs), offering a potentially more efficient and flexible alternative to traditional physics-based and data-driven methods. Traditional physics-based models, while reliable, require extensive domain expertise and can be computationally demanding. Data-driven methods, on the other hand, often struggle with extrapolation and require large datasets. VAEs offer a promising middle ground, learning the underlying data distribution and enabling the generation of new, realistic data samples.
The researchers trained VAEs on experimental data collected from two variants of a fully electric SUV, differing in maximum torque delivery and drivetrain configuration. The data, consisting of jerk signals derived from acceleration measurements, was preprocessed using Short-Time Fourier Transform (STFT) to create spectrograms, which were then used to train the VAEs. Three VAE architectures were investigated: unconditional VAE, conditional VAE (CVAE), and Gaussian Mixture Model CVAE (GMM-CVAE). The unconditional VAE learned the general data distribution without explicit input parameters, while the CVAEs conditioned the generation process on specific torque demands, making them more suitable for practical simulation scenarios.
Performance evaluation showed that the unconditional VAE achieved the lowest reconstruction errors, with an MSE of 0.3551 and MAE of 0.4203. It also outperformed the conditional variants in structural similarity (SSIM = 0.6082) and noise handling (PSNR = 16.7421 dB, SNR = 9.6040 dB). However, the CVAEs offered the crucial advantage of generating jerk signals conditioned on specific torque demands, a key requirement for realistic simulations. The GMM-CVAE did not offer substantial performance gains over the standard CVAE. New jerk signals were generated by sampling from the latent space using the reparametrization trick:
z = μφ + σφ(x) ⊙ ε
where z represents the latent variable, μφ and σφ are the mean and standard deviation from the encoder, and ε is a noise vector.
A comparison with a physics-based hardware model and a hybrid model (combining a simplified physics model with data-driven correction) further highlighted the VAEs' potential. While the hybrid model achieved slightly better accuracy, its complexity, relying on a 7-million parameter U-Net, made it significantly more data-intensive. The VAEs, especially the CVAEs, provide a compelling balance between accuracy and efficiency, particularly valuable in data-scarce scenarios. They offer a promising tool for augmenting drivetrain datasets and simulating complex driving scenarios.
This newsletter highlights a convergence of advanced techniques in signal processing, communication, and sensing. The exploration of large language models like RadioLLM for cognitive radio tasks represents a significant leap towards more intelligent and adaptable wireless communication systems. The research on brainwave biometrics, leveraging large-scale datasets, demonstrates the growing potential of this technology while acknowledging the challenges of real-world deployment. Finally, the application of variational autoencoders for drivetrain simulation showcases the power of generative AI in tackling complex engineering problems, particularly in data-limited scenarios. These advancements collectively pave the way for next-generation systems capable of handling the increasing complexities of our connected world.