HIPPOCAMPAL NEURAL SEQUENCES DIVERGE TO REFLECT DYNAMIC ENVIRONMENTAL GOALS WHILE GENERALIZING GLOBAL TASK STRUCTURES ACROSS CONTEXTS

The hippocampus is often described as the brain's locus for navigation, but its rich neural mixed selectivity indicates a broader role in memory and flexible cognition. We analyzed the geometry of hippocampal representations in the presence of dynamic behavioral goals, studying how changing task demands and sensation reshape the place code. Using an immersive virtual reality system, we trained mice to obtain rewards by locating a visual cue that could appear at different locations in the environment across trials. By recording large populations of CA1 pyramidal neurons using 2-photon calcium imaging, we found that the presence or absence of cues at each location drove unique patterns of remapping that decorrelated CA1 place codes up to hundreds of centimeters post-reward. Interleaving trials in a second, identically-sized virtual environment with no cues immediately elicited an orthogonal place code, despite the lack of explicit task in that environment. We then transferred the same task cues into the second context, and observed that mice immediately collected rewards at the cued locations, performing zero-shot inference. Neural decoding revealed that the mice formed a generalizable representation of the task structure that abstracted every possible cue-position combination across contexts, while preserving the ability to decode global context identity. This geometry collapsed in mice where the task cue was unrewarded in the second context. These data evidence the development of structured schemas in the hippocampus that can be flexibly deployed across contexts, and demonstrate that invariant place codes are largely absent in dynamic environments.