Today’s cloud-centric Large Language Model (LLM) inference can raise privacy concerns, as raw user data is sent over public networks. Split LLM inference offers a promising alternative by handling sensitive stages, such as prefill and input/output decoding, locally at the edge while offloading heavy intermediate decoding layers to the cloud. This approach could enhance privacy, yet it introduces a key challenge: how to transfer large intermediate vectors with ultra-low latency and high reliability over dynamic networks. We explore a potential transmission framework combining (i) dynamic congestion control to adapt flow and reduce queueing/loss, and (ii) adaptive FEC and quantization to balance redundancy with compression. Preliminary results indicate that this approach can achieve lower per-token latency and higher throughput than TCP/QUIC, enabling privacy-preserving LLM services that remain highly responsive and aligned with real-time user experience demands.

With the rapid development in the field of the Internet of Things (IoT), location information of people has become an indispensable part of achieving intelligent and digital management. Ultra-wideband (UWB) technology is widely used in indoor active positioning due to the characteristics of its pulse signals, which can achieve high-precision positioning within a range of a few centimeters. However, in device-free localization (DFL), the UWB tomography localization technique is highly disturbed by artifacts and requires a priori information about the number of known people to solve for the location of the target. In order to overcome this challenge, a training-free method is proposed in this paper to suppress radio tomographic imaging (RTI) artifacts in UWB systems. This method uses a single anchor L-shaped movement (LSM) to optimize the link layout and utilizes an improved elliptical model for precise linkpixel mapping. Finally, image reconstruction is achieved through guided filtering technology, effectively suppressing artifact interference. Experiments show that the method can achieve high-accuracy localization of multiple targets based on unknown a priori information about the number of people using received signal strength (RSS). When the number of positioning targets is 4, the quantity detection accuracy can reach 97%, and the system positioning error is less than 0.19 m.

This paper presents the theoretical analysis and experimental investigation of actively controlled manipulation of a magnetic microbead using quadrupole magnetic tweezers. Bead dynamics, magnetic actuation, and visual measurement are analyzed. A feedback control law is developed and implemented to stabilize and steer the motion of the magnetic microbead. It is developed in two steps. First, an inverse model, which is associated with a lumped-parameter analytical force model, is derived to enable feedback linearization. Second, linear controllers are designed to achieve motion stabilization and manipulation of the magnetic microbead. A proportional-gain controller along with feedback linearization is implemented to establish a stable trapping of the magnetic bead to facilitate system calibration. Experiments are then performed to validate the derived inverse force model and theoretical analysis. In addition, a minimum-variance controller is designed and employed to reduce the variance of the bead's Brownian motion. The control performance in terms of variance reduction, nanostepping, and large-range steering is then experimentally demonstrated.