This shift is especially important in the era of large AI models. As large language models, vision-language models, and other foundation models become increasingly integrated into communication systems, the information being transmitted is no longer limited to raw symbols or packets. It may involve semantic features, embeddings, natural language descriptions, multimodal representations, or task-specific intermediate outputs. This creates new opportunities for efficient wireless communication, but also introduces new challenges in robustness, reliability, and deployment.
In our recent work, Distributionally Robust Wireless Semantic Communication with Large AI Models, accepted by the IEEE Journal on Selected Areas in Communications, we study this problem from the perspective of robust semantic transmission under wireless uncertainty and distribution shifts.
Traditional wireless communication systems are usually designed around bit-level reliability. Metrics such as bit error rate, signal-to-noise ratio, and channel capacity are central to system design. These metrics remain fundamental, but they do not fully capture the requirements of AI-driven communication systems.
For many intelligent applications, the final goal is not merely to reconstruct the exact transmitted signal. Instead, the goal may be to support a downstream task, such as classification, retrieval, reasoning, decision making, or natural language understanding. In these settings, different parts of the transmitted information may have different semantic importance. Some errors may be harmless, while others may significantly change the meaning or degrade the downstream task.
This motivates a semantic view of communication. A semantic communication system should preserve the information that matters most for the intended task, while using wireless resources efficiently.
Large AI models provide a powerful tool for extracting, compressing, and interpreting semantic information. Compared with traditional handcrafted feature extractors, large models can represent richer and more flexible semantic structures. They can also support more complex tasks involving language, vision, multimodal understanding, and reasoning.
However, using large AI models in wireless semantic communication is not straightforward. These models are often computationally expensive, sensitive to distribution shifts, and difficult to deploy over resource-constrained wireless or edge environments. In practice, the data distribution at deployment time may differ from the training distribution. Channel conditions may also fluctuate due to mobility, interference, fading, and bandwidth limitations.
As a result, semantic communication systems based on large AI models must be designed not only for average-case performance, but also for robustness under uncertain and changing environments.
A key challenge in wireless semantic communication is distribution shift. A model trained under one data distribution or channel condition may perform poorly when deployed under another. This is particularly problematic for semantic systems, because semantic errors can be difficult to detect using conventional bit-level metrics.
Distributionally robust optimization provides one way to address this issue. Instead of optimizing performance only on a fixed empirical distribution, distributionally robust methods consider a family of possible distributions around the observed data. The goal is to learn a system that performs well even under plausible shifts from the training environment.
In the context of wireless semantic communication, this perspective is valuable because both the semantic source and the wireless channel may change. A robust system should therefore account for uncertainty in the transmitted content, the model representation, and the communication environment.
Another important motivation behind this research direction is edge intelligence. Many future AI-driven wireless applications will run on devices or networks with limited computation, memory, bandwidth, and energy. This includes mobile devices, sensors, vehicles, robots, and distributed edge servers.
Large AI models are powerful, but their direct deployment on edge devices is challenging. This creates a tension between semantic intelligence and system efficiency. On one hand, we want to use expressive models to capture useful semantic information. On the other hand, we need communication and computation mechanisms that remain feasible under practical resource constraints.
This tension is one of the reasons I am interested in the intersection of semantic communication, federated learning, and efficient AI systems. These areas all ask a similar question: how can we make intelligent systems reliable and useful when computation, communication, and data are distributed and constrained?
Working on wireless semantic communication with large AI models helped me better understand the connection between communication systems and modern AI. The problem is not only about applying neural networks to wireless transmission. It is also about rethinking what should be transmitted, what should be preserved, and how robustness should be measured when the receiver is performing an intelligent task.
This direction also raises several open questions that I find particularly interesting:
These questions are closely related to my broader research interests in federated learning, semantic communication, efficient AI systems, and large AI model deployment over constrained networks.
Wireless semantic communication with large AI models is an exciting research area because it connects communication theory, machine learning, optimization, and edge intelligence. As AI systems become more distributed and communication networks become more intelligent, preserving meaning rather than only preserving bits will become increasingly important.
For me, this work represents an important step in understanding how robust and efficient AI systems can be built over real-world wireless and edge networks.
]]>I plan to use this space for short technical notes on federated learning, semantic communication, parameter-efficient tuning, and the engineering practices that make research experiments reproducible.
]]>