Han Zhang @hanzhang.bsky.social · 09/09/2026Here, across thousands of naturalistic scenes, I tested whether VLMs show similar behavior, and indeed they do. Furthermore, model predictions aligned more closely with human fixations under matching goals than under mismatched goals. 100
Han Zhang @hanzhang.bsky.social · 09/09/2026This alignment also shows in decoder-layer attention and intermediate thinking traces. Search thinking is semantically close to the target (a [bottle] could be near the table, under the table, behind the dishes), whereas free-view thinking is close to visual prominence. 110
Han Zhang @hanzhang.bsky.social · 09/09/2026So, broad multimodal learning may acquire scene priors that support flexible attentional selection, even without direct gaze supervision. VLMs could become scalable models for where people look under different task goals. Feedback welcome! 000
