StreamGaze, accepted at NeurIPS 2026, benchmarks whether multimodal models can use human gaze

Daeun Lee and Adobe Research introduce StreamGaze, the first benchmark testing how well multimodal language models use human gaze signals for temporal reasoning in video.