The Semantic Communications Panel, hosted by the ATIS Next G Alliance (NGA) Technology Roadmap Working Group in early summer 2026, convened experts from industry and academia to discuss the future of semantic communications (SemCom) in the 6G era. The panel built on contributions spanning the last year from InterDigital, Ericsson, Nokia, Virginia Tech, Ohio State University, and Futurewei. The topic challenges one of the most basic assumptions of communications — that the job of a network is to move bits reliably — and instead asks what information matters to the receiver’s task. This shift in design philosophy can unlock greater efficiency and reliability over the next decade.

Defining Semantic Communications

Panelists broadly agreed that SemCom aims to convey meaning, or task-relevant information, rather than to reconstruct every transmitted bit. Beyond that shared starting point, framings diverged in instructive ways. One information theorist view cautioned against over-reliance on Weaver’s classic three-level model, arguing that much of what SemCom needs can be built on modern information theory. Others preferred to sidestep “meaning” altogether, noting that the term is human-centric and hard to pin down. Several panelists organized the problem around the receiver’s task and the importance of information to it — a framing that extends naturally to machines and physical AI. A related debate was whether SemCom requires joint optimization of source coding, channel coding, and inference; the panel treated this as one valid approach rather than a strict requirement.

Semantic Communication vs. Semantic Compression

A recurring theme was distinguishing SemCom from semantic compression. Professor Aylin Yener (Ohio State) offered the cleanest line: communication assumes a noisy channel between sender and receiver, whereas compression does not — so semantic communication spans the full path from information generation to the conveyance of meaning, with compression as one component. Compression alone doesn’t guarantee that what’s discarded is unimportant to the task. SemCom adds that layer — pairing compression with shared knowledge of importance and a defined task, and learning a model of the data that supports reasoning and new tasks at the receiver. Professor Walid Saad (Virginia Tech), who frames SemCom as a steppingstone toward artificial-general-intelligence (AGI)-native networks, positioned it as the “perception” piece of a world model — the point at which meaning is represented — while acknowledging that full generalizability remains an aspiration. In this view, semantic communications can also be seen as an AI-native layer that enables cross-layer contextual awareness and task-oriented operation within a network.

Use Cases and a Roadmap

Panelists converged on a roadmap that begins with constrained, machine-oriented traffic and expands toward immersive, human-centric applications:

1. Within-network data, such as channel state information feedback — conveying only what the channel estimate is used for (e.g., downlink precoding) rather than reconstructing the channel itself.
2. Machine-to-machine traffic with clear, frequently repeated tasks, where reduced data volume is a welcome by-product of task-focused design. This also includes communication among AI agents powering end-user machines.
3. Physical AI, including cooperative perception in vehicular settings and industrial digital twins — task-driven use cases with large data volumes.
4. Immersive and holographic communications in the longer term, potentially sending a compact description that a generative model reconstructs at the receiver with far less bandwidth.

Whether sheer data growth will make SemCom unavoidable drew measured answers: the panel saw the changing nature of AI-driven traffic — and the need to differentiate traffic by reliability and latency — as stronger motivations than volume alone. Value was seen across the ecosystem, from hyperscalers and device vendors to industrial users and operators (notably through integrated sensing and communication, or ISAC), with semantic compression likely to arrive first because it is the simplest step and builds on techniques already in wide use.

Key Technology Enablers

The panel split the enablers into a “6G for semantic and semantic for 6G” view. On the application side, Professor Saad emphasized world models and causal reasoning as the AI grounding for a semantic layer, alongside large language models, vision-action models, and object classification for extracting semantics. On the network side, Dr. Harish Viswanathan (Nokia) called for new interfaces to carry semantic context from applications into the network, evolving today’s quality-of-service APIs toward richer metadata and enabling differentiated handling such as unequal error protection and importance-aware retransmission.

Dr. Konstantinos Vandikas (Ericsson) cautioned that the transformer-based architecture underpinning world models scales quadratically with sequence length, motivating work on efficiency — key-value cache, distillation — and on meaning-fidelity metrics that can double as optimization objectives. The panel agreed that, at least near-term, much can be achieved within existing networking layers without fully disruptive, end-to-end redesigns.

Challenges and Open Questions

• Defining and measuring meaning. “Meaning” still lacks a formal, task-independent definition, and metrics remain the sticking point. Professor Yener noted candidly that after 15 years, she still lacks a semantic metric she is happy with. That said, meaning will have to be defined based on the fundamental data structure in the information being conveyed. The group nonetheless coalesced around task success rate (for example, tasks correctly completed per unit time, or the minimum data and latency needed to hit a target) as a workable basis for comparison.
• The network’s role and coexistence. Because general-purpose networks must carry all kinds of traffic, joint optimization is limited; the open question is what importance information the network is given and what it does with it, pointing toward a hybrid that preserves application/network separation while opening interfaces between them.
• Standardization, datasets, and governance. Panelists debated whether a sufficiently general, learnable semantic layer needs standardizing at all, agreed that community benchmarks are valuable but hard to build for real-time, end-to-end systems, and flagged a new security surface — conveying meaning could let an adversary induce incorrect decisions, so physical-AI applications will need governance and task-specific limits on permissible actions.

Next Steps for the Next G Alliance

Asked for one concrete NGA action to advance the work taking place in this area, panelists aligned on a pragmatic sequence:

• First, agree on terminology
• Then prioritize a few forward-looking use cases — ones likely to matter as SemCom matures
• Sketch a step-by-step evolution path for each use case
• Treat rigorous semantic metrics as an important but mid-term goal

Understanding the role of semantics, and of each entity in the end-to-end chain, for those use cases should in turn reveal the architectures and interfaces required.

Final Thoughts

This NGA panel discussion highlighted both the promise and the challenges of semantic communications. Panelists broadly agreed that semantic compression will come first, that early value is most plausible in machine-to-machine and physical-AI settings, and that immersive, human-centric applications are a longer-term prospect. They also stressed that meaning remains hard to define and to measure, that the network’s role is still unsettled, and that governance and security deserve early attention. As the industry moves toward AI-native 6G, the panel’s guidance was to start small and concrete — aligning on terminology and a few high-value use cases — and to let that work reveal the metrics, architectures, and interfaces that semantic communications will ultimately require.