Reading minds from eye movement, even when data drops out

Cognitive load assessment from eye-tracking could make autonomous systems safer, but blinks and sensor failures create gaps in the data, and existing models are computationally expensive. MambaGaze uses a technique that explicitly encodes missing data alongside raw features, then applies bidirectional Mamba-2 (a linear-complexity temporal model) to capture long-range patterns. On two standard datasets, it reaches 73–77% accuracy, beating transformers and CNNs by 4–12 points, while running at 43–68 frames per second on edge devices like NVIDIA Jetson with minimal power draw—making it practical for wearable deployment.