Semantic sensing (SemS) aims to acquire task‑relevant information instead of reconstructing the full physical scene. Existing SemS approaches are typically open‑loop: sensing configurations and observation schedules are fixed before inference and cannot adapt to evolving task‑level evidence.
We introduce Agentic Semantic Sensing, a closed‑loop framework tailored for AI‑enabled radio access networks (AI‑RAN). The framework actively controls sensing within a set of communication‑feasible profiles, allowing the sensing process to react to task demands in real time.
At its core lies a profile‑conditioned causal Transformer that continuously updates the semantic belief from streaming observations. To avoid re‑processing the entire history when profiles change, a key‑value cache is employed, enabling efficient incremental state updates based only on new observations.
In parallel, a semantic utility network estimates the task‑level benefit of acquiring the next observation block under each feasible profile while explicitly accounting for sensing cost. The network outputs continuation utilities that jointly support two decisions: selecting the next sensing profile and determining whether to perform a semantic early exit, i.e., terminating sensing early when evidence is sufficient to save resources.
From a theoretical perspective, the expected semantic gain can be related to conditional mutual information, offering a value‑of‑information interpretation for continued online sensing.
Experiments on the Widar3.0 dataset with six emulated sensing profiles show that, compared with the full‑sequence High configuration, the resource‑efficient Agentic setting reduces normalized cumulative sensing cost by 25.33% while achieving 85.79% Macro‑F1. Introducing semantic early exit at the same utility checkpoint further cuts cost by 12.35%, with only a 0.97‑point drop in Macro‑F1.
Review