Eye-tracking AI for cognitive load hits a new benchmark — and fits inside a drone controller's power budget.
Researchers behind MambaGaze, detailed in arXiv paper 2605.22775, built a bidirectional Mamba-2 model to infer how mentally taxed a person is from eye-gaze signals. The system tackles two persistent problems: gaze data goes missing constantly from blinks and tracker dropouts, and capturing how mental state evolves over time is computationally expensive. Their fix is XMD encoding, which stuffs observation masks and log-scaled time-deltas alongside the raw gaze features so the model always knows when data is absent and for how long. On two datasets — CLARE and CL-Drive — evaluated under leave-one-subject-out protocol, MambaGaze hit 77.1% and 69.4% accuracy respectively, and, critically, achieved the highest average macro-F1 across all ten compared models at 55.3%.
The macro-F1 figure is the one that actually matters here. Accuracy looks clean until you remember that cognitive load datasets are class-imbalanced; macro-F1 penalizes models that coast on the majority class. Besting ten competitors on that metric, not just raw accuracy, is a harder claim to wave away. The edge deployment results add a second layer of credibility: 27-36 frames per second on NVIDIA Jetson Orin hardware drawing under 6.6 watts puts real-time inference within reach of embedded systems, not just server racks.
Driver vigilance monitoring and flight deck assistance are the use cases the authors name, and both involve exactly the intermittent data and long time horizons that killed earlier approaches. The 5-20 percentage point macro-F1 gain from combining all three XMD streams — versus any single channel — suggests the missing-data modeling is doing real work, not just padding the feature vector. Whether these numbers hold outside lab datasets with consenting subjects is the next question nobody has answered yet.