AI is set to transform the telecommunications industry—and ATIS is leading the way in advancing AI Network Applications. This work focuses on enabling networks to develop a semantic understanding of the physical wireless environment, allowing them to adapt intelligently to real-time conditions and user intent. These innovations lay the foundation for truly cognitive, AI-native wireless systems. One key area of progress is the development of the Wireless Physical-Layer Foundation Model (WPFM).
The Challenge
-
- AI in wireless is currently task-specific and fragmented, with limited applicability across the wireless domain. For example, in indoor localization applications, tasks such as Non-Line-of-Sight (NLOS) detection, error correction, or beamforming solutions are typically built as bespoke models. Similarly, in spectrum management, tasks like wireless technology recognition, modulation recognition, and interference detection are typically built on separate AI models and do not share any knowledge.
-
- AI models often fail to generalize across real-world conditions due to limited training data.
-
- Wireless environments are dynamic, but current AI solutions lack adaptability.
-
- Large Language Models (LLMs) understand language semantics; no equivalent currently exists for physical wireless signals.
💡 The Solution – WPFM
A WPFM is a foundation model built specifically for the wireless physical layer.
-
- A foundation model is (pre-)trained on a vast amount of unlabeled data and can be adapted (finetuning) to a variety of downstream tasks, even those involving different types of data.
-
- Similarly, LLMs can be adapted to many different tasks, but work specifically with written language.
-
- With the WPFM, the goal is to develop a foundation model capable of understanding physical-layer (RF) modalities, typically represented as IQ-based time series, through techniques such as pattern recognition, representation learning, embeddings, and generative AI.
Beyond the above factors, the WPFM aims to enable semantic association between RF signals and their wireless context (e.g., labels/text), moving beyond traditional pattern recognition. Similar to cross-modality encoders trained to encode raw data using contrastive learning approaches, such as CLIP (Contrastive Language–Image Pre-Training), the WPFM’s goal is to maximize cross-modality similarity between the environment and RF signals. This approach creates meaningful connections between RF patterns and visual or textual representations, enhancing the interpretability and contextual awareness of wireless signal analysis.
As such, WPFMs are designed to bridge fundamental gaps between RF signals and human-interpretable concepts through cross-modal learning techniques and symbolic reasoning. WPFMs can provide embeddings or symbolic structures to a neuro-symbolic pipeline. This not only allows the WPFM to classify the RF status but also explains why it was classified. This enables the creation of verifiable semantic descriptions of wireless network conditions, sensing applications, human behavior, and the simultaneous classification and localization of signals, among other applications, from wireless data. This granular understanding allows systems to comprehend signal relationships and contexts rather than merely identifying isolated patterns. In future work, the WPFM aims to be integrated with NeSy, a neuro-symbolic cognitive assistant.
In short, the WPFM is designed for multi-task learning, cross-environment generalization, real-time adaptability, and semantic RF understanding.
Key Advantages
| ✅ Unified Model | 🧠 Semantic RF Understanding |
| 🔄 Scalable & Adaptable | 🚀 Accelerates Innovation |
| 🤝 Hybrid AI + Domain Expertise | 📊 Multi-task Capable |
📈 Why WPFM Matters for the Wireless Future
-
- Zero-touch automation: Supports self-optimizing, intent-driven networks.
-
- 6G readiness: A foundational enabler of cognitive and autonomous future networks.
-
- Robust performance: Learns across spectrum bands, geographies, and topologies.
-
- AI-native networking: Embeds intelligence directly into the wireless fabric.
-
- A foundational shift for scalable, cognitive, and adaptive wireless networks.
WPFM for Wireless Networks: Help Shape Its Direction with Your Operator or Vendor Insight
The WPFM represents a new class of AI explicitly built for the physical dynamics of wireless networks. But its true value depends on how it addresses your toughest challenges.
We’re seeking input from network operators, infrastructure vendors, and others across the wireless ecosystem to help shape the direction of this innovation.
We invite you to reflect:
-
- What wireless applications or use cases are you exploring where current AI tools fall short?
-
- Are your models struggling to generalize across different environments, devices, or spectrum bands?
-
- Do you face challenges with real-time adaptability, spectrum awareness, or efficient edge deployment?
-
- Would semantic understanding of RF signals help your systems operate more autonomously or intelligently?
-
- Are there other high-priority areas—technical, operational, or architectural—where a foundation model like WPFM could provide meaningful support?
WPFM is designed to be a flexible foundation, capable of supporting multi-task learning, real-time responsiveness, and integration with existing wireless architectures. For applications such as spectrum management, network control, or RF sensing, the WPFM provides a framework for incorporating semantic intelligence directly at the physical layer, enabling cross-task generalization and adaptive behavior within dynamic wireless environments.
Get Involved
The ATIS AI Network Applications (ANA) group is collaborating with imec and Ghent University to refine the requirements for the WPFM. To learn more about the WPFM initiative and how to get involved, please contact Rich Moran at rmoran@atis.org.
