Semantic communication asks whether a shared model of meaning lets a sender transmit less

Faithful symbols and conveyed meaning

The distinction is one of objective, not technique.

Classical information theory treats communication as the problem of reproducing a message at a destination with acceptable fidelity, and sets meaning outside the engineering problem. That choice is what made the field tractable: capacity, noise and error correction can be defined without asking what a message is about. A channel is good if the symbols arrive intact.

Semantic communication reopens the part that was set aside. It accepts that the symbols may arrive imperfectly, or not at all, provided the receiver reconstructs the intended meaning — which means sender and receiver must agree on something about how that meaning is represented. Once that agreement exists, it is a resource: knowledge both ends hold, which does not have to be paid for again on every transmission.

The three research directions

A contemporary survey identifies three leading directions, which is a useful map of where the work actually is.

• Theory-of-mind-based. Work that models what the receiver is likely to know, intend, or misunderstand, and shapes the transmission around that estimate rather than the raw content.
• Generative-AI-driven. Work in which a generative model reconstructs content at the receiver from a compact representation rather than reproducing the original signal. An ICASSP special session overview frames the challenge as extracting semantic information from complex content and regenerating semantically consistent data, robust to channel corruption.
• Deep joint source-channel coding. Work that learns the source encoding and the channel encoding together, in one model, instead of treating compression and transmission as separate stages.

What the three share is the premise that the receiver is not a passive destination. It is a system with its own model, and that model carries part of the load.

Why shared knowledge reduces what must be sent

Because a message only has to carry what the receiver cannot already derive.

If both ends hold a compatible model of the domain, the sender can transmit a reference, a correction, or a task rather than a full description, and the receiver regenerates the rest. This is the same economy that makes context engineering work: a well-assembled context spends tokens on the difference between what the model knows and what the task requires. The context engineering page develops that argument for model windows, and token savings covers what it costs. Semantic communication makes the same case for a physical channel, where the shared model is what shrinks the message.

There is an obvious dependency. The economy holds only while the shared model is actually shared. A receiver whose model has drifted will reconstruct a different meaning from the same transmission, and do so confidently. Shared knowledge buys the reduction; maintaining it is what the reduction costs.

What is speculative

Most of it, in the sense that matters to anyone building a system today.

The surveys describe a research programme: theory, architectures and simulation-based evaluation, largely in wireless and network settings. They do not describe a capability a team can adopt this quarter. Reported gains should be read against their experimental setup — what channel model, what baseline, what task — because the results are conditional on those choices in a way headline numbers do not convey. This page makes no performance claim, and neither should a procurement decision.

What is worth taking is the framing rather than the numbers. If a receiver has a model of meaning, the sender can send less — and if that model has drifted, the sender has sent less and communicated the wrong thing. That is the trade the agentic side of this site makes. See agentic AI for what happens when inferences become actions, and A2A for how agents that must agree on meaning describe themselves.