Khalifa University RF AI Foundation Model: Pioneering RF-GPT for Radio Frequency Spectrograms

The Breakthrough: Khalifa University's RF-GPT Ushers in a New Era for Wireless AI

Khalifa University of Science and Technology in Abu Dhabi has achieved a groundbreaking milestone in artificial intelligence applied to telecommunications. On February 17, 2026, the university's Digital Future Institute announced RF-GPT, recognized as the world's first radio-frequency (RF) language model, specifically designed as a foundation model for processing radio frequency spectrograms.6059 This innovation transforms complex wireless signals into understandable visual representations, enabling AI systems to analyze, interpret, and reason about them using natural language. For researchers, engineers, and students in the United Arab Emirates' thriving higher education sector, RF-GPT represents not just a technical leap but a strategic advancement in positioning the UAE as a global leader in 6G and AI-native networks.

Traditional approaches to RF signal processing rely on specialized, task-specific algorithms that struggle with the dynamic, multifaceted nature of modern wireless environments. RF-GPT changes this paradigm by integrating RF data directly into multimodal large language models (LLMs), much like how vision-language models handle images alongside text. This development stems from the Khalifa University 6G Research Center (6GRC), directed by Professor Mérouane Debbah, whose team includes experts like Hang Zou, Yu Tian, Bohao Wang, Lina Bariah, Samson Lasaulce, and Chongwen Huang.61

Understanding Radio Frequency Spectrograms and the Challenges They Pose

Radio frequency spectrograms are two-dimensional visual maps that depict the distribution of signal power across time and frequency, derived from complex in-phase and quadrature (IQ) waveforms captured in wireless communications. These spectrograms reveal critical details such as modulation schemes, bandwidth allocation, signal overlaps, and interference patterns—essential for technologies like 5G New Radio (NR), Long-Term Evolution (LTE), Wi-Fi (WLAN), and satellite communications (DVB-S2).61

Processing these spectrograms has historically required bespoke deep learning models for each task, such as modulation recognition or anomaly detection. These models demand vast labeled datasets, extensive computational resources, and retraining for new scenarios, limiting scalability in real-world deployments. In the UAE, where rapid 5G expansion and 6G preparations are underway, such inefficiencies hinder spectrum management and network optimization. RF-GPT addresses this by treating spectrograms as images, leveraging pretrained vision encoders to generate 'RF tokens' that an LLM can process alongside textual instructions.60

• Spectrograms capture time-frequency structures, making patterns like resource blocks in 5G NR visually discernible.
• Challenges include overlapping signals, varying sampling rates, noise, and device-specific distortions.
• Foundation models offer generalization via pretraining on diverse data, enabling zero-shot performance on unseen tasks.

Inside RF-GPT: Architecture and Step-by-Step Processing Pipeline

The RF-GPT architecture is elegantly simple yet profoundly effective, building on established vision-language models without requiring architectural overhauls. Here's how it processes RF data step-by-step:

1. Capture and Transform: Raw IQ waveforms from standards-compliant generators (e.g., MATLAB toolboxes for 5G, LTE) are converted to spectrograms using Short-Time Fourier Transform (STFT) with a Blackman window, 512-point FFT, and 512-sample hop, yielding 512x512 pixel grayscale images normalized to dB scale.61
2. Visual Encoding: A pretrained Vision Transformer (ViT from Qwen2.5-VL) divides the spectrogram into 14x14 patches, producing 1,369 RF tokens via multi-head self-attention and MLP layers.
3. Token Adaptation: A lightweight linear projection maps RF tokens into the LLM's embedding space.
4. Reasoning and Generation: Tokens feed into a decoder-only Transformer LLM (Qwen2.5-VL 3B or 7B parameters), which autoregressively generates natural language descriptions, classifications, or structured JSON outputs conditioned on user prompts like 'Identify modulations and overlaps in this spectrum.'

Training involves supervised fine-tuning (SFT) on synthetic data using AdamW optimizer, cosine decay, and mixed-precision on 8 NVIDIA H200 GPUs—completing in mere hours over 3 epochs.61 This efficiency democratizes advanced RF AI for UAE universities and industries.

The Power of Synthetic Data: Building RFDraw Without Manual Labeling

A key innovation is the RFDraw dataset: 12,000 unique RF scenes encompassing six technologies (5G NR, LTE, UMTS, WLAN, DVB-S2, Bluetooth), augmented with channel effects like fading and noise across SNR ranges. Waveform generators ensure standards compliance via rejection sampling.

Captioning hierarchies (summary to signal-level details) feed a text-only LLM (GPT-OSS-120B) to synthesize 625,000 instruction-answer pairs, eliminating costly annotations. This scalable approach yields diverse tasks: modulation classification, overlap detection, user counting, and parameter extraction.61

For UAE researchers, this method lowers barriers to RF AI development, fostering innovation in higher education labs.

Benchmark Results: RF-GPT Dominates RF Intelligence Tasks

RF-GPT was rigorously evaluated on five benchmarks, outperforming general-purpose VLMs (e.g., GPT-4V, Qwen2.5-VL raw) which fail without RF priors, and matching/exceeding task-specific CNN/ViT models with far less data.61

Robustness tests show resilience to impairments like carrier frequency offset and phase noise. These results validate RF-GPT's generalization, crucial for UAE's dynamic telecom landscape.Read the full paper.

Real-World Applications: From Spectrum Management to Defense

RF-GPT's versatility spans telecom operators troubleshooting interference, regulators ensuring compliance, and defense sectors achieving spectrum awareness in contested environments. In 6G, it enables autonomous networks querying 'Is this 5G-WLAN overlap compliant?' for real-time optimization.60

• Telecom: Anomaly detection, resource allocation in integrated sensing-communications (ISAC).
• Industrial: Multimodal RF sensing for localization and scene understanding.
• UAE Security: Unauthorized signal detection amid growing drone/UAV traffic.

Imagine a UAE telco engineer prompting RF-GPT for 5G NR subcarrier spacing from a spectrogram—deployable via edge computing.

UAE's Strategic Push: Khalifa University and the 6G Vision

Khalifa University's prowess aligns with UAE's National AI Strategy 2031, emphasizing sovereign AI and 6G leadership. The 6GRC has won ITU challenges and partnered with GSMA, du, Nokia for TelecomGPT.11 RF-GPT bolsters this, training on Arabic-inclusive benchmarks for regional relevance.

In UAE higher education, KU's output—86% Q1 journals, 575% top-10% papers—elevates global rankings. This fosters PhD programs in AI-telecom, attracting talent via scholarships.Explore UAE scholarships.

Stakeholder Perspectives: Voices from the Research Frontier

Professor Debbah states, 'RF-GPT represents a turning point... allowing networks to manage themselves.' Lina Bariah highlights scalability for Arabic telecom LLMs. Industry echoes: GSMA Foundry eyes RF-GPT for benchmarks.

UAE academics praise KU's role in bridging academia-industry, per recent forums. Challenges like data privacy in RF AI are addressed via federated learning research.

Future Outlook: Scaling RF AI for 6G and Beyond

Future enhancements include real-world datasets, larger models, and integration with agentic AI for closed-loop control. For UAE, RF-GPT accelerates 6G trials, smart cities, and defense autonomy. Ethical considerations—bias mitigation, explainability—are prioritized.59

Opportunities abound: Aspiring researchers can pursue research assistant jobs at KU, leveraging RF-GPT for theses.

Photo by Arin Melikyan on Unsplash

Career Implications in UAE Higher Education

This breakthrough signals booming demand for AI-wireless experts. KU's programs in electrical engineering and AI produce graduates for etisalat, du, and startups. Craft your academic CV to join. Internships via 6GRC offer hands-on RF AI experience.

• Skills: Python, PyTorch, RF simulation (MATLAB), LLMs.
• Roles: RF ML Engineer, 6G Researcher, Spectrum Analyst.

Check UAE university jobs for openings.