fMRI与CNN-LSTM建模揭示灵活决策中共享关系图的几何重格式化

Adaptive behavior requires balancing generalization from past experience with the flexible selection of goal-relevant information. Cognitive maps provide a structured basis for this balance, yet how a stable relational map supports task-appropriate decision coordinates remains unclear. Combining fMRI with CNN-LSTM modeling, we show that a shared two-dimensional social map coexists with localized, goal-dependent geometric reformatting. A hippocampal-orbitofrontal/medial prefrontal network preserves a reusable 2D relational structure, whereas a frontoparietal control network expresses goal-aligned 1D decision projections. Crucially, these readouts show systematic geometric shaping along distinct functional axes. Representational stretching along a "congruent axis" was associated with redundant structure across dimensions, whereas goal-directed warping along a "task-relevant axis" selectively emphasized currently relevant information during decision conflict. This localized geometric reformatting was accompanied by dynamic hippocampus-insula-precuneus interactions. As similar patterns emerged spontaneously in recurrent networks, our findings support geometric reformatting as a complementary readout strategy through which a stable multidimensional relational map can be flexibly adapted to current decision demands.

中文翻译还没有生成，先显示原文。