Anna Mynick¹, Adam Steel², Caroline Robertson¹; ¹Dartmouth College, ²University of Illinois Urbana-Champaign

Memory facilitates seamless perception in immersive environments (Mynick et al., 2025), setting up predictions of upcoming scene views that lie beyond our immediate field of view. Yet, how this process of memory-based scene prediction occurs in the brain is unknown. Here, we asked how scene predictions and their consequences are represented across the scene processing hierarchy: (i) the occipital pole, (ii) the scene perception areas (“SPAs”: PPA, OPA), (iii) the place memory areas (“PMAs”: VPMA, LPMA) (Steel et al., 2021), which sit immediately anterior to each of the SPAs, (iv) the medial place area (MPA), which is both mnemonic and perceptual (Steel et al., 2025), and (v) the hippocampus. Participants (N=16) first learned an immersive scene in head-mounted VR. On each trial of the subsequent fMRI task, we presented a sequence of three images depicting 30° views from the studied scene. During the third view ("outcome"), participants reported whether the sequence represented a spatially accurate sweep across the environment. On valid trials (57%), all three scene views were consecutive and adjacent, allowing the outcome to be predicted. On invalid and omission trials, the first two views were consecutive and adjacent, but the predicted outcome view was replaced by either an incorrect view (invalid trials), or a blank grey square (omission trials), violating the prediction. We observed robust evidence for predictions and violation responses across the scene processing hierarchy, with different patterns for perceptual (SPAs) and mnemonic areas (PMAs). Specifically, mnemonic areas (hippocampus, MPA, and the PMAs) showed representations of predicted-but-omitted scene views (all p<.05), while the SPAs and occipital pole did not. Further, the MPA and the PMAs showed evidence for prediction errors, with greater univariate activity for invalid versus valid trials (p<.01). Taken together, these results suggest that the hippocampus and PMAs support memory-based scene predictions in real-world places.