The development of large language models (LLMs) has promoted a transformation of human–computer interaction (HCI) models and has attracted the attention of scholars to the evaluation of personality traits of LLMs. As an important interface for the HCI and human–machine interface (HMI) in the future, the intelligent cockpit has become one of LLM’s most important application scenarios. When in-vehicle intelligent systems based on in-vehicle LLMs begin to become human assistants or even partners, it has become important to study the “personality” of in-vehicle LLMs. Referring to the relevant research on personality traits of LLMs, this study selected the psychological scales Big Five Inventory-2 (BFI-2), Myers–Briggs Type Indicator (MBTI), and Short Dark Triad (SD-3) to establish a personality traits evaluation framework for in-vehicle LLMs. Then, we used this framework to evaluate the personality of three in-vehicle LLMs. The results showed that psychological scales can be used to measure the personality traits of in-vehicle LLMs. In-vehicle LLMs showed commonalities in extroversion, agreeableness, conscientiousness, and action patterns, yet differences in openness, perception, decision-making, information acquisition methods, and psychopathy. According to the results, we established anthropomorphic personality personas of different in-vehicle LLMs. This study represents a novel attempt to evaluate the personalities of in-vehicle LLMs. The experimental results deepen our understanding of in-vehicle LLMs and contribute to the further exploration of personalized fine-tuning of in-vehicle LLMs and the improvement in the user experience of the automobile in the future.

Conventional eye trackers are difficult to deploy in fashion-runway settings because they require head stabilization, wearable devices, close-range cameras, or infrared illumination. This study develops a far-field, non-contact eye-tracking framework and prototype system for measuring target-directed visual attention in runway-viewing settings using standard cameras. The framework integrates eye-region normalization, 3D gaze estimation, individualized homography-based point-of-regard mapping onto a scene plane, adaptive filtering, DBSCAN-based outlier removal, and kernel density estimation. In Experiment 1, 16 participants completed a controlled calibration-grid benchmark with the proposed system and EyeLink 1000 Plus. Compared with the EyeLink 1000 Plus, the proposed system exhibited higher positional error ($A_{\mathrm{pos}}$: $2.50\pm 0.25°$ vs. $1.74\pm 0.55°$) and a higher value on the processed-output spatial-precision measure ($A_{\mathrm{spa}}$: $2.31\pm 0.24°$ vs. $1.02\pm 0.37°$). These results characterize end-to-end localization under the calibrated benchmark geometry and provide a controlled reference for the subsequent AOI-level evaluation. In Experiment 2, an independent sample of 16 participants completed an instructed right chest targeting task at 2 m, 4 m and 6 m. AOI-retained sample percentages were 80.4%, 97.3%, and 96.0%, respectively, and gaze-density maps concentrated within the predefined upper-body target region. These findings support the feasibility of target-directed AOI-level localization at far-field distances without chin-rest stabilization under the tested simulated runway-observation protocol.