Haichun Zhang

BACKGROUND Large language models (LLMs) can perform well on medical question-answering and diagnostic benchmarks, but their output remains sensitive to the instruction used at inference time. For diagnostic tasks, this sensitivity may affect whether a model reaches a definitive final diagnosis. OBJECTIVE This study evaluated how inference-time diagnostic instructions affect diagnostic reliability across models and text-based diagnostic datasets. METHODS We conducted a cross-sectional evaluation of 18 diagnostic instructions across 11 large language models and 3 diagnostic datasets. Physicians reviewed responses while blinded to model identity and explicit instruction labels. Responses were classified as completely correct, partially correct, completely incorrect, or uncertain. Each model by dataset combination formed one paired evaluation block, giving 33 blocks. DirectAsk was the reference condition. RESULTS Completely correct responses differed across the 18 diagnostic instructions, with Friedman χ²=118.44, P<.001. Relative to DirectAsk, Knowledge-Base and Knowledge-ClinicalStudy had the largest mean increases at +2.03 percentage points, followed by Role-Expert at +1.97 percentage points. After Holm adjustment, CoT and Role-Expert remained higher than DirectAsk. Style-Skeptical was lower than DirectAsk by -8.73 percentage points after Holm adjustment, and MultiExpert-Discussion changed by -3.14 percentage points. Style-Skeptical and MultiExpert-Discussion increased uncertain responses by +9.33 and +7.29 percentage points, respectively. The minimum block-level changes were -26.0 percentage points for Style-Skeptical, -18.0 for Stepwise-MultiTurn, and -17.0 for MultiExpert-Discussion. CONCLUSIONS Diagnostic instructions beyond DirectAsk did not show a large consistent performance advantage. Some instructions produced larger negative changes and more non-definitive final responses when the task required one diagnosis. Diagnostic instructions should be validated for the intended model and diagnostic task.

Deep Neural Networks (DNNs) perform intensive matrix multiplications but can tolerate inaccurate intermediate results to some degree. This makes them a perfect target for energy reduction by approximate computing. However, current research in this direction requires DNNs redesign and does not provide the flexibility for users to trade accuracy for energy saving. In this paper, we propose a runtime reconfigurable approximate floating-point multiplier and present details of its hardware implementation. The flexible computation precision is provided by our error correction module, which is controlled by reconfigurable clock signals. The circuit design solves the glitch and metastability problems. The proposed approximate multiplier with three precision levels is evaluated on Synopsys design compiler and Xilinx FPGA platforms. Experimental results demonstrate the advantages of our approach in terms of speed, hardware overhead, and power consumption, while ensuring a controllable accuracy loss for DNNs inferences.

NAND flash memory is widely used in communications, commercial servers, and cloud storage devices with a series of advantages such as high density, low cost, high speed, anti-magnetic, and anti-vibration. However, the reliability is increasingly getting worse while process improvements and technological advancements have brought higher storage densities to NAND flash memory. The degradation of reliability not only reduces the lifetime of the NAND flash memory but also causes the devices to be replaced prematurely based on the nominal value far below the minimum actual value, resulting in a great waste of lifetime. Using machine learning algorithms to accurately predict endurance levels can optimize wear-leveling strategies and warn bad memory blocks, which is of great significance for effectively extending the lifetime of NAND flash memory devices and avoiding serious losses caused by sudden failures. In this work, a multi-class endurance prediction scheme based on the SVM algorithm is proposed, which can predict the remaining P-E cycle level and the raw bit error level after various P-E cycles. Feature analysis based on endurance data is used to determine the basic elements of the model. Based on the error features, we present a variety of targeted optimization strategies, such as extracting the numerical features closely related to the endurance, and reducing the noise interference of transient faults through short-term repeated operations. Besides a high-parallel flash test platform supporting multiple protocols, a feature preprocessing module is constructed based on the ZYNQ-7030 chip. The pipelined module of SVM decision model can complete a single prediction within 37 us.

The Internet of Things (IoT) is an industry-recognized next intelligent life solution that increases the level of comfort, efficiency, and automation for citizens through numerous sensors, smart devices, and cloud stations connected physically. As an important application scenario of IoT, the Internet of Vehicles (IoV) plays an extremely critical role in the intelligent transportation field. In fact, the In-Vehicle Network of smart vehicles that are recognized as the core roles in intelligent transportation is currently the Controller Area Network (CAN). However, the In-Vehicle CAN bus protocol has several vulnerabilities without any encryption, authentication, or integrity checking, which severely threatens the safety of drivers and passengers. Once malicious attackers hack the vehicular gateway and obtain the access right of the CAN, they may control the vehicle based on the vulnerabilities of the CAN bus protocol. Given the severe security risk of CAN, we proposed the CANsec, a practical In-Vehicle CAN security evaluation tool that simulates malicious attacks according to major attack models to evaluate the security risk of the In-Vehicle CAN. We also show a usage case of the CANsec without knowing any information from the vehicle manufacturer.