Abstract
No two individuals’ gaze patterns during naturalistic viewing are identical: we each prioritize attending to different features when exploring a real-world environment. Recent studies have shown that gaze patterns are not only explained by attention to semantic content (e.g., phone, car) but also abstract conceptual information (e.g., for sale; being repaired) that can be modeled using computational language models. Are there unique patterns in individuals’ gaze toward conceptual information as they attend in real-world environments, and are these patterns stable over time for an individual? Here, we tested whether gaze patterns in conceptual space can individuate viewers across testing days, thus reflecting attentional priorities that are trait-, rather than state-like. We measured participants’ (N = 29) visual attention while they explored real-world photospheres in VR, across two sessions (Day 1: N = 60; Day 2: N = 40). We characterized the conceptual information in each scene by decomposing photospheres into a set of overlapping tiles and obtaining context-informed verbal descriptions from online raters. Tile descriptions were then transformed into sentence-level embeddings using a computational language model (BERT), capable of capturing context-dependent conceptual meaning (e.g., river bank vs. financial bank). To characterize the conceptual information a participant prioritized across all photospheres, we fit a model between this conceptual feature space and gaze behavior for each participant and testing session. We assessed whether individual conceptual priorities are stable and individuating across sessions by calculating the correlation between feature weights modeled for each participant’s two sessions. We find that conceptual feature weights were stable (r = 0.72) and individuating across days: within-subject feature weight correlation was higher, on average, than between-subject correlation (t(28) = 2.4, p = 0.03). Our results suggest that individual gaze patterns reflect high-level conceptual information seeking and individuating, trait-like priorities within conceptual space that persist across viewing sessions.