<p>Comparison of model complexity and efficiency.</p>

AR/VR and other immersive technologies are creating dynamic, learner-centred, and engaging language-learning environments. In these ever-changing situations, judging someone’s language abilities is difficult. Managing multimodal learner inputs, understanding model predictions, and protecting user data across distributed systems are some of the most prominent challenges. This paper proposes TriNet-AQ, a federated, interpretable deep learning architecture for classifying English competency in AR/VR platforms. This technique addresses the difficulties raised. This work employs Quantum Sinusoidal Encoding (QSE), Triaxial Attention Fusion (TAF) for multimodal feature alignment, and Quantum Modulated Integration (QMI) to enhance context-aware learning by optimizing temporal representation. Hybrid Slime Gorilla Optimisation (HSGO) aids optimization. It accelerates convergence and improves performance and economy. TriNet-AQ provides decentralized training to many clients via federated learning, enhancing privacy and flexibility. TriNet-AQ outperforms classical, fuzzy, and hybrid baselines in real-world augmented and virtual reality instructional datasets. Its accuracy is 98.5%, AUC is 0.95, and EPES is 0.89. Even when it loses 3.5% accuracy on new data, it can generalize effectively. Another SHAP-based interpretability finding is the presence of obvious feature attributions and consistent relevance across users. Statistical analysis, including Cohen’s d = 0.89 (p < 0.001), confirms the model’s significance and reliability. TriNet-AQ provides robust, easy-to-understand, and private real-time, tailored language evaluation in next-generation immersive learning environments.

增强现实/虚拟现实（AR/VR）及其他沉浸式技术正在构建动态化、以学习者为中心且极具吸引力的语言学习环境。在这些持续演化的场景中，对学习者的语言能力进行评估颇具挑战。处理多模态学习者输入、解读模型预测结果，以及在分布式系统中保护用户数据，均为当前最突出的几类难题。
本研究提出TriNet-AQ，一种用于在AR/VR平台上开展英语能力分级的联邦学习（Federated Learning）可解释深度学习架构，该技术可有效解决上述挑战。本研究采用量子正弦编码（Quantum Sinusoidal Encoding，QSE）、用于多模态特征对齐的三轴注意力融合（Triaxial Attention Fusion，TAF），以及量子调制集成（Quantum Modulated Integration，QMI），通过优化时序表征以增强情境感知学习能力。混合黏菌大猩猩优化算法（Hybrid Slime Gorilla Optimisation，HSGO）可助力优化流程，加速模型收敛并提升性能与资源经济性。
TriNet-AQ通过联邦学习实现面向多客户端的分布式训练，进一步强化了隐私性与灵活性。
在真实增强现实与虚拟现实教学数据集上，TriNet-AQ的表现优于经典、模糊逻辑及混合基线模型：其准确率达98.5%，曲线下面积（Area Under Curve，AUC）为0.95，期望预测误差平方（Expected Prediction Error Squared，EPES）为0.89。即便在新数据集上出现3.5%的精度损失，该模型仍可实现出色的泛化性能。
基于夏普利可加解释（SHAP）的可解释性分析结果进一步显示，该模型具备清晰的特征归因能力，且在不同用户群体间呈现出一致的相关性权重。包含科恩d值（Cohen’s d）=0.89（p<0.001）在内的统计分析结果，证实了该模型的显著性与可靠性。
TriNet-AQ可为下一代沉浸式学习环境提供鲁棒性强、易于解读且兼顾隐私保护的实时个性化语言评估服务。