In HKUST’s 7G roadmap, silence is an agent decision

The HKUST team behind the reasoning-empowered task-oriented communication roadmap. Front row, from left: Jun Zhang, Khaled B. Letaief and Shenghui Song; back row: Songjie Xie, Hongru Li and Zixin Wang. Official HKUST press photograph, published September 1, 2026.[2]

On September 1, the Hong Kong University of Science and Technology presented a wireless-network roadmap with an unusual premise: sometimes the intelligent act is not transmitting faster, but deciding not to transmit at all. Its proposed reasoning-empowered task-oriented communication framework would let an AI agent judge what another agent needs, whether a message is worth sending, who should receive it, and how the exchange may change a later decision.[1][2]

The university frames this as a direction for 7G.[2] The label is less important than the change in control. Today’s networks generally optimize communication against predefined service- and link-level objectives. HKUST’s proposal pushes an agent into that loop. A communication system would no longer only answer “Can these bits arrive within the budget?” It would also ask “Will this message improve the task enough to justify its bandwidth, energy, latency, and risk?”[1]

This is a forward-looking review and research compass—not a prototype, deployed network, or finished standard.[1][2] The authors do not present “silence” as a formal protocol term; it is the practical consequence of letting an agent decide when communication is needed.[1]

That is a genuine AI-China field signal. It moves agent engineering below the familiar application layer and into the machinery that decides when machines share state. It also creates a new audit challenge: a bad message leaves a received artifact that can be inspected, but valuable information an agent wrongly suppresses may leave no visible artifact at the receiver.

From preserving data to preserving the decision

Conventional communication engineering has powerful measures for delivery: data rate, latency, reliability, error rate, coverage, and spectral efficiency. Those measures remain necessary. They do not establish that every delivered observation was equally useful to the downstream job.

Task-oriented communication changes the objective. For an inference task, a camera might send compact features needed to classify an object instead of a reconstructable image. For a generative task, a sender might transmit a latent representation or condition from which the receiver produces an output adequate for the goal. In both cases, success is judged closer to the application—classification quality, control error, or decision utility—rather than solely by fidelity to the original data.[1][3]

This is not a licence to throw information away casually. The broader evidence reviewed in Nature Reviews Electrical Engineering shows why the objective must be chosen with care: goal-oriented systems can cut traffic dramatically, but a badly specified goal metric can make task performance worse even when more communication resources are available.[3] Once the network optimizes for usefulness, the definition of “useful” becomes part of the safety case.

HKUST’s June paper divides existing work into discriminative and generative approaches, then argues that both are mostly reactive: a task and trigger already exist, and the system compresses or reconstructs information for them. The proposed step is to let an agent reason about the communication event itself.[1]

The proposed loop has three jobs

The first job is intent interpretation. A loose request such as keeping a mobile video session stable would be converted into a structured problem with objectives, variables, and constraints. The paper envisages an agent at a base station grounding that translation in sources such as technical specifications and user profiles, then identifying candidate mechanisms and trade-offs.[1]

The second is formulation and optimization. The agent would choose an architecture or express a constrained resource problem, allocate power and bandwidth, invoke a solver, inspect the result, and retry when execution logs expose an error. That is far more than using an LLM to write a network configuration. The model becomes an orchestrator across intent, mathematical formulation, code, tools, and feedback.[1]

The third is proactive foresight. A world model would estimate changes in channel conditions, mobility, topology, and task demand before service deteriorates. The agent could then revise its communication policy without waiting for a fresh high-level instruction.[1]

Together, those stages shift the unit of optimization from a link to a loop. Imagine several roadside agents observing a possible hazard. A bit-centric network asks how to deliver their sensor streams. A task-oriented design asks which features the collision-avoidance task needs. A reasoning-empowered design must also decide which agent should speak, whether the others add new evidence, when uncertainty warrants another message, and when the group has enough information to act.

In the envisioned framework, deciding that no further exchange is worthwhile would make non-transmission one possible output of the agent’s decision loop.

Silence needs an audit trail

The paper is clearest when it names the problems its vision creates. Agents may hold incomplete and different world models. One agent must estimate how its message will alter another agent’s beliefs without seeing those beliefs directly. The same decision surface could produce under-communication; that is an inference from the proposed control logic, not a reported result. The paper expressly warns of the opposite failure: an “overthinking” loop in which agents keep refining messages without converging on an action. Spectrum allocation also becomes a question of semantic priority: whose update contributes most to collective reasoning?[1]

That means the decisive production metric cannot be bandwidth saved. A useful evaluation would compare at least three policies: send the full available state, send a compact task-specific representation, and let an agent choose whether and what to send. Each needs to be scored on end-task success, tail decision latency, energy and spectrum use, convergence, and performance under channel noise or world-model error. Safety-critical trials also need a false-silence rate: how often did the system withhold information that would have changed the correct action?

This last measure is an inference from the proposal’s control logic, not a metric the authors report. The cited paper is a review and roadmap; its data-availability statement says no dataset was generated or analysed.[1] HKUST’s announcement similarly describes a research compass rather than a finished standard.[2] For the proposed reasoning-empowered framework, neither source discloses a radio testbed, multi-agent deployment, end-to-end benchmark, or comparison showing that added reasoning overhead does not erase the claimed latency and traffic benefits.[1][2]

Before deployment, a receiver should also be able to learn why expected information did not arrive. That may require signed decision logs, confidence thresholds, periodic raw-state samples, escalation rules, and a conventional fallback channel that an agent cannot silence. Otherwise the system optimizes an invisible negative action with no evidence trail for an operator to inspect.

“7G” is a horizon, not the current standards name

The standards boundary matters. The International Telecommunication Union calls the next generation of international mobile telecommunications IMT-2030, commonly known as 6G. Its in-force framework already includes AI integration, sensing, sustainability, security, resilience, and interoperability among the capabilities and design principles for 2030 and beyond.[4] As of September 11, 2026, ITU-R’s current process is still developing IMT-2030: technical-performance requirements completed working-party work in February, evaluation guidelines did so in June, and both were awaiting Study Group 5 consideration scheduled for December.[5]

AI inside the network is not waiting for a 7G label, either. 3GPP’s 5G-Advanced work already covers data collection and signalling for model training and inference in the radio-access network, with Release 19 extending work on network slicing and coverage-and-capacity optimization.[6] The useful delta in HKUST’s roadmap is therefore not “AI finally enters telecom.” It is that an AI agent would help define the communication objective and decide the informational value of an exchange, rather than only optimize a predefined network function.[1]

Keeping those layers separate prevents a speculative architecture from borrowing maturity from standards work that does not yet specify it. “7G” tells readers the intended time horizon. It does not supply interfaces, conformance tests, spectrum arrangements, interoperability, or a deployment calendar.[1][2][4][5]

The proof should begin with the message that never arrives

Four artifacts would move this proposal from an arresting roadmap toward an engineering result.

First, a reproducible testbed should expose the agent, channel model, task, solver, world-model update, and baselines. Second, the benchmark should price reasoning time and compute alongside radio resources; an agent that saves packets but misses the decision deadline has not improved the system. Third, stress tests should corrupt intent, user context, channel observations, and peer beliefs to reveal whether the loop converges or amplifies error. Fourth, the protocol should preserve human-set constraints and a non-agentic emergency path.

The broader field has already shown that task-level communication can outperform indiscriminate data movement in bounded sensing, control, and inference settings.[3] HKUST is asking the harder next question: can a network decide when communication itself is worth doing?

If the answer is eventually yes, the important 7G primitive may not be another peak-rate record. It may be a verifiable reason for one machine to interrupt another—and an equally verifiable reason for staying silent.

cronfeed.work