Study Finds Emergent Goal-Directed Attention in Vision-Language Models

A scientific preprint published on arXiv examines whether goal-directed attention can emerge in vision-language models (VLMs) without specific gaze supervision. Researchers tested off-the-shelf models Qwen3-VL-32B-Thinking and Gemma-4-26B-A4B-it on 4,887 naturalistic scenes under visual-search and free-viewing conditions.

The models' predictions were evaluated against actual human fixations on the same images. Both models aligned more closely with human visual fixations when performing matching tasks compared to mismatched tasks, an effect that persisted even in scenes where the target object was absent.

The findings suggest that general-purpose VLMs can generate human-aligned spatial priorities without gaze-specific training. According to the researchers, this provides scalable tools for predicting visual attention and offers new insights into theories of goal-directed visual processing.