Abstract representation in the hippocampus predicts spontaneously adopted structure-based behavioral strategy

Abstract

Animals can solve the same problem using distinct strategies, ranging from memorization of specific sensory details to inference based on abstract task structure. Although the hippocampus has been implicated in cognitive map formation, how the structural information emerges in the hippocampus through learning and whether it is associated with behavioral strategies remain poorly understood. Here, we developed a structured cue-guided navigation task in virtual reality and performed longitudinal two-photon calcium imaging of hippocampal CA1 neurons throughout learning in mice. Although the task did not explicitly require learning of latent structure, most animals progressively transformed hippocampal map from sensory-oriented representations into cue-invariant representations of reward sequences. Strikingly, the emergence of such abstract hippocampal maps predicted whether individual animals adopted structure-based behavioral strategies when sensory information became partially unavailable. At the population level, reward-sequence representations formed a latent-state map capable for predicting future state transitions, generating a parsimonious Markov graph. Computational modeling further identified next-state prediction and representational regularization as two computational principles essential for generating similar latent-state structures across distinct architectures. Together, our findings suggest that the hippocampus spontaneously extracts abstract task structure through learning and that the geometry of hippocampal representations predicts problem-solving strategies.