CONVERGENT POPULATION DYNAMICS AND DIVERGENT MANIFOLD GEOMETRIES EMERGE DURING VOLITIONAL LEARNING IN THE HIPPOCAMPUS AND MOTOR CORTEX

Volitional control of neural activity via brain–computer interfaces (BCIs) provides a causal framework to test how circuits reshape population dynamics during learning. However, most BCI studies focus on primary cortex, leaving open whether associative regions can support volitional control. Here we tested whether hippocampal CA3 can be recruited for BCI control and compared learning-related reorganization to primary motor cortex (M1). We developed a two-photon calcium imaging BCI in which head-fixed mice learned to bidirectionally modulate selected neuronal ensembles in either M1 or CA3 to obtain reward. Across days, we quantified learning curves and population sparsity, trial-aligned dynamics, and low-dimensional geometry using state-space analyses and nonparametric manifold labeling of reward-aligned trajectories. Mice acquired robust control in both regions, showing that CA3 supports volitional modulation despite distinct circuit architecture. Learning increased sparsity and strengthened alignment of ensemble trajectories with task demands in both M1 and CA3. However, temporal structure of adaptation diverged: M1 showed prominent pre-reward excitation consistent with preparatory drive, whereas CA3 exhibited strong post-reward inhibition suggestive of recurrent stabilization and reset-like dynamics. Low-dimensional manifold analysis further revealed region-specific geometric transformations, with increasing separability of movement- versus stay-preceded trajectories in M1 but convergence in CA3. Both regions increased recurrence and occupancy of reward-manifold states across sessions, with CA3 showing higher traversal directionality consistent with stabilization of encoded task structure. Together, these results reveal complementary and divergent circuit mechanisms and indicate that distinct architectures can converge on a shared functional outcome for volitional control of population activity during BCI learning.