AI News

Wi-Fi channel state information (CSI) supports device-free sensing, including human activity recognition, but sensing models can become unreliable when deployment users or environments differ from training conditions. Existing approaches generally address this as an offline model-design problem, such as learning stronger representations or applying one adaptation method across the target domain. The CSI-Agent paper instead treats adaptation as a deployment-time decision problem, motivated by scarce labeled target data and the fact that different classes may fail differently under the same domain shift. CSI-Agent is an evidence-seeking LLM agent that does not process raw CSI or make sample-level predictions. It summarizes target-domain behavior into sensing-grounded, class-level evidence, establishes a target-adaptive default from complementary CSI views, and uses an LLM planner to decide whether each class should keep that default or invoke a specialized action. Deterministic verification and bounded execution are used to limit unreliable interventions. The authors evaluate the system on four public datasets across five cross-domain splits covering device, user, environment, and compositional shifts. In 1-shot adaptation, CSI-Agent achieves the best target-domain performance on every split and raises average Macro-F1 by about 16% over the strongest baseline.