GazeEval-VLM: Gaze-Guided Evaluation for Vision-Language Models

Yingpeng Ma. Manuscript in preparation, 2026.

An independent research project building a lightweight, model-agnostic evaluation framework for Vision-Language Models, using human gaze as an external cognitive signal. Using Qwen2.5-VL-3B as the base model and the VQA-MHUG and ChartGaze datasets, the project studies how gaze-guided visual interventions, such as gaze crops, gaze highlighting, blurred backgrounds, and multi-view gaze-guided inputs, affect model performance on visual and chart question answering tasks. It also introduces GazeVLM-Lite, a lightweight inference method that uses gaze as visual evidence to improve model performance, and a gaze efficiency score to weigh performance gains against the added attention cost.