"IRIS: Intent Resolution via Inference-time Saccades for Open-Ended VQA in Large Vision-Language Models Dataset"

"This dataset accompanies IRIS (Intent Resolution via Inference-time Saccades) and contains eye-tracking and VQA trial data collected from a controlled user study of 500 unique image-question pairs spanning ambiguous and unambiguous visual queries. Each trial includes a full participant fixation sequence with normalized screen coordinates and timestamps, the participant's verbally posed question, a ground-truth answer, and responses from 11 state-of-the-art Vision-Language Models (VLMs) evaluated under three conditions: without eye data, with temporal filtering only, and with spatiotemporal filtering. Timing metadata captures the onset and offset of each participant's spoken question, enabling precise alignment of gaze behavior with verbal query initiation. The dataset is intended to support research on gaze-guided disambiguation in open-ended VQA and serves as a benchmark for evaluating the integration of real-time eye movement data into Large Vision-Language Models."