With the rapid development of large language models (LLMs), their capabilities have expanded from pure text generation to wide applications including image recognition, audio understanding, and more. However, the physical world contains a vast range of sensing data beyond vision and audio—such as temperature, air pressure, acceleration, voltage, and electromagnetic signal strength. From a broader perspective, our team explores whether LLMs can understand and process these types of data and thus understand the physical world.

We designed two experiments to evaluate LLMs’ ability to handle physical signals. The results show that LLMs can infer users’ real-world activities and location contexts from textualized smartphone sensor signals, and can also detect heartbeats by analyzing digitized electrocardiogram (ECG) data. These findings reveal LLMs’ latent capability for interpreting the physical world. Building on this insight, we introduce a new concept—Penetrative AI—which refers to using LLMs’ embedded world knowledge to understand and process signals from widely deployed IoT sensors or actuators, thereby enabling perception and decision-making tasks in cyber-physical systems (CPS).

We further extend Penetrative AI from the individual device level to the network level, proposing the broader concept of a Penetrative IoT, which integrates LLMs with IoT for distributed sensing, holistic reasoning, and coordinated deployment. Compared with traditional paradigms, Penetrative IoT leverages the general knowledge embedded in LLMs to provide richer support for cyber-physical systems, representing a new direction and promising potential for LLMs-empowered IoT.