Maren Wehrheim^1,2, Kohitij Kar¹; ¹York University, Department of Biology and Centre for Vision Research, Centre for Integrative and Applied Neuroscience, Toronto, Canada, ²Mila - Quebec AI institute, Montreal, Canada

Causal evidence shows that the ventrolateral prefrontal cortex (vlPFC) is necessary for core object recognition: vlPFC inactivation degrades behavioral performance and IT-based object decodes, with larger impairments for late-solved, challenging images. Yet, we still lack a computational account of how vlPFC participates in the ventral-stream hierarchy. Does vlPFC act as a downstream readout of IT, or does it supply recurrent, image-specific signals that help IT refine object representations, particularly for challenging images? These hypotheses make distinct predictions. A pure readout model holds that vlPFC should (i) produce object decodes more aligned with behavior than IT, and (ii) contribute no additional task-relevant structure beyond what is already represented in IT. A recurrent-enhancer model instead predicts that (i) vlPFC will be less behaviorally aligned than IT, (ii) vlPFC activity will explain IT variance in an object-solution-time (OST)-specific manner, and (iii) removing vlPFC-derived components from IT should impair decoding, especially for late-solved images. To adjudicate between these models, we recorded from IT (437 neurons) and vlPFC (234 neurons) in two macaques performing object recognition on 1,320 images. As expected, early-solved images (OST 90–110 ms) yielded rapid, high-accuracy IT decodes, whereas late-solved images (>180 ms) yielded slower, weaker IT evidence. Contrary to the readout model, vlPFC decodes were not more behaviorally aligned than IT, and vlPFC contributed substantial image-dependent explained variance (~10% of IT responses), markedly larger for late-OST images. Residual analyses showed that removing vlPFC-predictable components from IT responses reduced decoding performance, with strongest impairments for late-solved images, demonstrating that vlPFC provides behaviorally relevant structure not present in feedforward IT. Noise-corrected dimensionality analyses further showed that vlPFC constrains IT representations toward task-relevant manifolds for late-solved images. Together, these results position vlPFC not as a downstream readout, but as a selective recurrent partner that shapes IT population structure during challenging object recognition.

Acknowledgements: MW is funded by Connected Minds Postdoctoral Fellowship (supported by CFREF). KK was supported by the Canada Research Chair Program (CRC-2021-00326), Canada First Research Excellence Funds (VISTA Program), & the National Sciences and Engineering Research Council of Canada (NSERC, RGPIN-2024-06223).