arXiv cs.LG→ original

Scientists Find Anhedonia-Like Symptoms in Vision-Language Models Similar to Human Depression

Researchers tested whether human-like qualities in vision-language models extend to reward evaluation using clinical anhedonia tests—a depression symptom linked to the nucleus accumbens (NAc). They identified model units responsible for reward anticipation and demonstrated that perturbing them shifts the model toward lower-effort, lower-reward options while preserving core capabilities.

AI-processed from arXiv cs.LG; edited by Hamidun News
Scientists Find Anhedonia-Like Symptoms in Vision-Language Models Similar to Human Depression
Source: arXiv cs.LG. Collage: Hamidun News.
◐ Listen to article

Researchers published a paper on arXiv examining whether the ability of vision-language models to capture human behavior extends to reward assessment — using clinical tests originally developed to diagnose anhedonia and motivational impairments in major depressive disorder.

How researchers found the model's "reward center"

In humans, anhedonia is often linked to dysregulation of the nucleus accumbens (NAc) and the broader dopaminergic reward system of the brain. Neuroimaging can localize these impairments, but directly proving a causal relationship between NAc activity and specific behavioral symptoms remains difficult even in humans. The authors applied this logic to vision-language models: using methods from neuroscience, they functionally identified in the models units that anticipate reward — an analog of NAc — and then tested their causal role through targeted perturbations, that is, artificial interventions in the operation of precisely these units.

What the experiments showed

Perturbation of NAc-like units produces a behavioral effect in the model that mirrors anhedonia in humans: in tasks requiring selection 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 does not lose basic functionality: if you remove the reward-based choice element itself, the model shows normal, baseline results — meaning the deficit is specific to reward assessment and anticipation, not a general loss of capability. The result also correlates with clinical anhedonia scales, including DARS and MAP-SR.

What this means

The work demonstrates that circuits functionally similar to the human brain's reward system can be detected and deliberately modified in vision-language models. For AI interpretability, this opens a path to targeted diagnosis of model "behavioral" failures, and for neuroscience — provides another tool for testing hypotheses about anhedonia mechanisms that are difficult to test directly in humans.

Frequently asked questions

What is anhedonia?

The inability to experience pleasure and reduced motivation — one of the key symptoms of major depressive disorder. In humans, this state is associated with dysfunction of the nucleus accumbens and the dopaminergic reward system of the brain.

Does the model lose functionality after perturbation?

No. If you remove the reward-based choice element from the task, the model shows normal baseline results — the failure manifests precisely in decisions that require comparing effort and reward.

ZK
Hamidun News
AI news without noise. Daily editorial selection from 50+ sources. A product by Zhemal Khamidun, Head of AI at Alpina Digital.

Want to stop reading about AI and start using it?

AI News is a curated feed of AI/tech news. Hamidun Academy teaches you to use AI systematically in your work.

What do you think?
Loading comments…