AI models can suffer 'anhedonia' after targeted neuron-like perturbations

Researchers Honarmand, Mahdipour Aghabagher, and Schrimpf asked whether vision-language models (VLMs) align with human cognition in reward valuation. Using a mechanistic framework based on clinical tests for anhedonia and motivational deficits in major depressive disorder, they functionally identified reward-anticipatory units in VLMs analogous to the human Nucleus Accumbens (NAc) and its dopaminergic reward system. When they perturbed these NAc-selective units, the model exhibited behavioral effects strikingly similar to human anhedonia: it systematically shifted toward low-effort, low-reward options in effort-based decision-making tasks. Crucially, this was not a general performance drop—when reward-based choice was removed, the perturbed model maintained baseline accuracy, proving the deficit is specific to reward valuation and anticipation.

The induced vulnerability aligned with clinical anhedonia and motivation scales, including the DARS (Dimensional Anhedonia Rating Scale) and MAP-SR (Motivation and Pleasure Scale-Self-Report). These results reveal that reward valuation circuits in AI models can parallel those in humans, establishing a causal link between specific neural units and behavioral symptoms that has been difficult to achieve in human neuroimaging. This opens the door to using VLMs as in silico models for studying psychiatric conditions like depression, potentially accelerating the development of targeted therapies.