Researchers find a vision-language model analogue of human anhedonia in depression

Researchers have published a paper on arXiv testing whether the ability of vision-language models to capture human behavior extends to reward evaluation. They used clinical tests originally designed to diagnose anhedonia and motivational deficits in major depressive disorder.

How the researchers looked for a model's "reward center"

In humans, anhedonia is often linked to dysregulation of the nucleus accumbens (NAc) and the brain's broader dopamine reward system. Neuroimaging localizes these abnormalities, but directly proving a causal link between NAc activity and specific behavioral symptoms is still difficult, even in humans. The authors carried this logic over to vision-language models: using methods from neuroscience, they functionally identified units in the models that anticipate reward, an analogue of the NAc, and then tested their causal role through targeted perturbations, that is, artificial interventions in the operation of exactly those units.

What the experiments showed

Perturbing NAc-like units produces a behavioral effect in the model that closely mirrors human anhedonia. In tasks that require choosing between options with different effort-to-reward ratios (effort-based decision-making), the model begins to systematically prefer options with lower effort and lower reward. Importantly, the model's baseline performance is not lost: if the reward-based choice element is removed, the model produces its usual baseline results. The deficit is therefore specific to reward evaluation and anticipation, and is not a general loss of capability. The result also correlates with clinical anhedonia scales, including DARS and MAP-SR.

What this means

The work shows that circuits functionally similar to the human brain's reward system can be found and deliberately altered in vision-language models. For AI interpretability, this opens a path to pinpoint diagnosis of "behavioral" failures in models. For neuroscience, it offers another tool for testing hypotheses about the mechanisms of anhedonia that are hard to test directly in humans.

Frequently asked questions

What is anhedonia?

It is the inability to experience pleasure, along with reduced motivation, and is one of the key symptoms of major depressive disorder. In humans, the condition is linked to dysfunction of the nucleus accumbens and the brain's dopamine reward system.

Does the model lose its capabilities after perturbation?

No. If the reward-based choice element is removed from the task, the model produces its usual baseline results. The failure shows up specifically in decisions that require weighing effort against reward.