Beyond Shannon: UCSD’s KDC Framework Shifts Wireless Paradigms from Bits to Token-Native Communication

For over seven decades, wireless systems have been built on Shannon's foundational bargain: deliver every bit as faithfully as possible. 5G, Wi-Fi, and every protocol in between optimize for bit-level fidelity. But in real wireless environments with packet loss, fading channels, and unpredictable interference, that insistence on exact reconstruction often means higher latency, wasted bandwidth, and brittle performance on the tasks that actually matter.

A UC San Diego research team led by PhD student @Xingyu Chen from CWC Director Professor @Xinyu Zhang's group is proposing a fundamentally different paradigm. Their framework, KDC (Knowledge-Driven Communication), replaces the bit-centric communication with the token, a semantic unit grounded in meaning rather than signal reconstruction. Instead of treating all bits as equally important, KDC leverages multimodal foundation models to extract a lightweight, task-specific operator for each data modality, then prioritizes the tokens that carry the most knowledge for the downstream task.

On the receiver side, the system pairs a pretrained shared knowledge base with dynamically updated context, enabling it to fill in gaps and recover task-relevant meaning even under heavy packet loss. The result isn't just compression. It's a communication system that reasons about what to send.

Semantic communication has been widely studied in recent years, with a growing body of work exploring learned codecs, joint source-channel coding, and task-oriented transmission. KDC builds on this momentum but pushes the boundary further. Rather than optimizing a single neural codec end-to-end, it introduces a modular, knowledge-driven architecture that generalizes across modalities and tasks, laying the foundation for agent-to-agent, token-native communication.

The team tested KDC across image, video, LiDAR, and radar data, running experiments over real 5G, Wi-Fi, and LoRa links. Under lossy, unstable channel conditions, KDC maintained significantly stronger task performance compared to conventional approaches, demonstrating that tokens, not bits, may be the right abstraction for the AI-native wireless stack.

The work was featured in a talk at @Samsung Research America’s SPARK Sessions and will be presented at #NSDI '26 on May 4th in Renton, WA. Learn more at the project website: https://bit2token.github.io/

As AI agents increasingly rely on real-time multimodal perception to act in the physical world, the communication layer will need to keep up. KDC points toward that future: a post-Shannon framework where wireless systems don't just move data, but move knowledge.