Pupil detection is a key technology in fields such as human–computer interaction, fatigue driving detection, and medical diagnosis. Existing pupil detection algorithms still face challenges in maintaining robustness under variable lighting conditions and occlusion scenarios. In this paper, we propose a novel pupil detection algorithm, ViMSA, based on the ViM model. This algorithm introduces weighted feature fusion, aiming to enable the model to adaptively learn the contribution of different feature patches to the pupil detection results; combines ViM with the MSA (multi-head self-attention) mechanism), aiming to integrate global features and improve the accuracy and robustness of pupil detection; and uses FFT (Fast Fourier Transform) to convert the time-domain vector outer product in MSA into a frequency–domain dot product, in order to reduce the computational complexity of the model and improve the detection efficiency of the model. ViMSA was trained and tested on nearly 135,000 pupil images from 30 different datasets, demonstrating exceptional generalization capability. The experimental results demonstrate that the proposed ViMSA achieves 99.6% detection accuracy at five pixels with an RMSE of 1.67 pixels and a processing speed exceeding 100 FPS, meeting real-time monitoring requirements for various applications including operation under variable and uneven lighting conditions, assistive technology (enabling communication with neuro-motor disorder patients through pupil recognition), computer gaming, and automotive industry applications (enhancing traffic safety by monitoring drivers’ cognitive states).

Differentiable Neural Architecture Search has emerged as a powerful paradigm for automated network design, yet it suffers from a fundamental optimization inconsistency problem: Architectures optimized under continuous relaxation often fail to maintain their performance after discretization. To address this challenge, we propose Fisher-DARTS—a Fisher information-driven differentiable NAS framework. The proposed method introduces three key innovations: (1) a Fisher information-based momentum update mechanism that guides architectural parameters toward statistically significant operations, aligning the search objective with discrete deployment; (2) a progressive three-region pruning strategy that adaptively eliminates redundant operations with low Fisher information, ensuring architectural compactness; and (3) a cell-weighted fusion module that preserves multi-scale features across stacked cells. Additionally, the search space is expanded by incorporating attention mechanisms to enhance feature representation capability. The proposed framework is generic and applicable to a wide range of vision tasks. To validate its effectiveness, we apply it to gaze estimation—a core technology in multimodal human–computer interaction. Experimental results on three public datasets, MPIIFaceGaze, RT-GENE, and ETH-XGaze, demonstrate that Fisher-DARTS achieves mean angular errors of 3.22°, 5.45°, and 4.12°, respectively, outperforming hand-designed networks and existing NAS-based gaze estimation models. These results validate the effectiveness of the proposed Fisher-driven NAS framework and its generalization capability across diverse scenarios.