Tsz Wai Bentley Lo¹ (tlo57@uwo.ca), Susheel Vijayraghavan¹, Lyle Muller², Julio Martinez-Trujillo^1,3; ¹Department of Physiology and Pharmacology University of Western Ontario, ²Department of Mathematics University of Western Ontario, ³Department of Psychiatry, University of Western Ontario

Visual working memory (vWM) is a fundamental cognitive function essential for goal-directed behaviour; however, it is characterized by strict capacity limits. While the lateral prefrontal cortex (LPFC) is critical for vWM maintenance, the computational mechanisms constraining capacity in New World primates remain understudied. We investigated these limits in the common marmoset (Callithrix jacchus), leveraging their lissencephalic cortex to implant 3D volumetric arrays (N-Form, Plexon) in LPFC areas 8/46. Uniquely, we utilized wireless electrophysiology to record neural activity while animals performed a touchscreen-based Delayed Non-Match to Position (DNMP) task under freely moving conditions. Behaviourally, performance declined significantly with increasing load, reflecting classic capacity limitations. We recorded from 994 neurons in 2 marmosets. We identified single neurons (33%) exhibiting spatially selective persistent activity during memory maintenance. A topographical analysis identified a distributed "salt-and-pepper" functional architecture in LPFC, distinct from the retinotopy observed in early visual areas. To investigate the computational basis of capacity limits, we compared coding models of population activity. Neural firing rates scaled sublinearly with load, consistent with a divisive normalization model, which was statistically preferred over Summation and Winner-Take-All models based on model comparison metrics (Friedman p < 0.05). Principal Component Analysis (PCA) revealed that population activity for multi-item trials occupied a low-dimensional subspace intermediate between single-item subspaces. This "geometric averaging" suggests that normalization compresses the representational space to maintain dynamic range stability. However, it reduces the signal-to-noise ratio for representations of individual items. These findings suggest that vWM capacity limits are governed by population-level normalization dynamics in the LPFC.