Classical information theory, pioneered by Claude Shannon in 1948, focuses on the "Level A" problem: how to transmit symbols accurately over a noisy channel. As Shuai Ma 0002 and Kai Niu observe, Shannon’s theory intentionally ignored the "meaning" of those symbols to create a mathematically tractable framework for bits. However, as 6G networks approach the theoretical limits of spectral efficiency, researchers like Angeliki Alexiou and Deniz Gunduz argue for a paradigm shift toward Semantic Communication (SemCom).
The core intuition is simple: if the receiver understands the context or the goal of a message, the transmitter does not need to send every bit. By transmitting "meaning" rather than raw data, we can break the "bandwidth wall" and achieve massive gains in efficiency.
The Mechanism of Synonymous Mapping
In traditional systems, every unique sequence of bits is treated as a distinct message. Semantic communication changes this by introducing synonymous mapping. As Zijian Liang and Ping Zhang explain, many different data patterns can carry the same meaning for a specific task.
By grouping these patterns into "synonymous sets," a system can compress information far more aggressively. Instead of coding for the statistical uncertainty of symbols (Shannon entropy), we code for the uncertainty of meaning (semantic entropy). This allows the transmitter to "intellicise" the data—pruning redundant or irrelevant details that do not contribute to the receiver's understanding or the task's success.
Semantic Joint Source Channel Coding
In the Shannon paradigm, source coding (compression) and channel coding (error correction) are often treated separately. Semantic communication collapses this distinction through Semantic Joint Source-Channel Coding (SJSCC).
Jincheng Dai and Kai Niu propose frameworks where the "codec" is a deep neural network trained to preserve meaning across a noisy link. Rather than trying to reconstruct a pixel-perfect image or a word-perfect text, the system prioritizes features that are critical for the receiver's task. For example, in a vehicular network, Z. Wang suggests that a "utility-oriented" approach would prioritize the location of a pedestrian over the texture of the asphalt, drastically reducing the required bit rate.
Establishing Theoretical Speed Limits
A major hurdle identified by Marwa Chafii and Hai-Long Qin is the lack of a formal mathematical foundation equivalent to Shannon's capacity formula. Without it, engineers cannot know the "speed limits" of semantic transfer.
Recent work by Chuan Huang 0001 addresses this by developing the Semantic Rate-Distortion Function (SRDF). This framework calculates the minimum bit rate required to transmit both the raw observations and their underlying semantic states. By using neural-network-based estimators, researchers can now quantify how much potential improvement remains for semantic systems compared to traditional bit-level codecs.
Goal-Oriented and Agentic Communication
Semantic communication naturally evolves into goal-oriented communication. In this view, the "value" of information is defined by its impact on a physical process or an autonomous task.
Cheng Feng and Kaibin Huang describe systems where radio resources are allocated based on "information utility"—how much a packet reduces control error or improves model accuracy in a Cyber-Physical System. Taking this further, Haixiao Gao proposes Agentic AI-Enhanced SemCom, where Large Language Models (LLMs) act as autonomous agents that manage the communication loop, understanding user intent to execute tasks with minimal data exchange.
Generative AI as the New Codec
The rise of Generative AI (GAI) provides a powerful toolkit for this transition. Hai-Long Qin and Guo Lu highlight Diffusion Models as a way to enable extreme data compression. In this setup, the transmitter sends only a sparse "semantic prompt" or a scene graph, and the receiver uses a generative model to synthesize the full content.
As Chen Zhu demonstrates with scene-graph-aided image transmission, this approach can achieve over 200x the throughput of traditional JPEG compression. The receiver isn't just decoding bits; it is "re-imagining" the sender's intent based on shared knowledge, a process Guangming Shi calls "first understand, then transmit."