An Efficient Multimodal Causal Attention Architecture for Real-Time Mobile Educational Content Interpretation

The increasing adoption of mobile learning requires efficient multimodal models capable of real-time educational content analysis under resource constraints. However, existing multimodal architectures often involve high inference latency and memory requirements, limiting their deployment on mobile and edge devices. This study proposes the Efficient Multimodal Causal Attention Architecture (EMCAA), integrating speech, lecture transcripts, presentation slides, and contextual learner interactions. EMCAA combines Temporal Directed Acyclic Graph (T-DAG) Causal Attention to prioritize temporally relevant dependencies, Cross-Modal Layer Fusion (CMLF) to integrate complementary multimodal representations, and Causal ContextAware Quantization (CCQ) to reduce computational and memory requirements through adaptive precision allocation. Experiments on the Multimodal Student Attention Dataset showed that EMCAA achieved 91.4% accuracy, 91.0% precision, 90.8% recall, and 90.9% F1-score. Compared with MC-MIFA, EMCAA improved accuracy by 1.2 percentage points while reducing inference latency from 57 to 43 ms (24.6%) and memory consumption from 401 to 329 MB (18.0%). Ablation analysis further demonstrated the contribution of the proposed components. These findings indicate the potential of EMCAA for efficient multimodal educational content processing in resource-constrained mobile and edge environments.