Researchers Find a Vision-Language Model Analog of Human Anhedonia in Depression

Researchers tested whether the human-like behavior of vision-language models extends to reward valuation. To do it, they used clinical tests for anhedonia, a symptom of depression linked to the brain's nucleus accumbens (NAc). They identified units inside the models that anticipate reward and showed that perturbing them shifts the model toward lower-effort, lower-reward options. The models' baseline capabilities remain intact.

AI-processed from arXiv cs.LG; edited by Hamidun News
Researchers Find a Vision-Language Model Analog of Human Anhedonia in Depression
Source: arXiv cs.LG. Collage: Hamidun News.
Listen to article

In a new paper on arXiv, researchers tested whether vision-language models' ability to capture human behavior extends to reward valuation. Their tools were clinical tests originally designed to diagnose anhedonia and motivational deficits in major depressive disorder.

How they found the model's "reward center"

In humans, anhedonia is often linked to dysregulation of the nucleus accumbens (NAc) and the brain's broader dopaminergic reward system. Neuroimaging can localize these disruptions, but directly proving a causal link between NAc activity and specific behavioral symptoms remains difficult, even in humans. The authors carried this logic over to vision-language models. Using methods borrowed from neuroscience, they functionally identified units in the models that anticipate reward, an analog of the NAc. They then tested the causal role of these units with targeted perturbations: artificial interventions applied to those units specifically.

What the experiments showed

Perturbing the NAc-like units produces a behavioral effect that closely mirrors anhedonia in humans. In effort-based decision-making tasks, where the model chooses between options with different trade-offs between effort and reward, it starts systematically preferring lower-effort, lower-reward options. Crucially, the model does not lose its general competence. When the reward-based choice element is removed, the model performs at its normal baseline. In other words, the deficit is specific to reward valuation and anticipation, not a general loss of capability. The results also align with clinical anhedonia scales, including DARS and MAP-SR.

What it means

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

FAQ

What is anhedonia?

Anhedonia is the inability to feel pleasure, combined with reduced motivation. It 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 dopaminergic reward system.

Does the model lose its general competence after perturbation?

No. When the reward-based choice element is removed from a task, the model performs at its normal baseline. The failure shows up specifically in decisions that require weighing effort against reward.

Key Questions Answered

What is anhedonia?

Anhedonia is the inability to feel pleasure, combined with reduced motivation. It 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 dopaminergic reward system.

Does the model lose its general competence after perturbation?

No. When the reward-based choice element is removed from a task, the model performs at its normal baseline. The failure shows up specifically in decisions that require weighing effort against reward.

How they found the model's "reward center"?

In humans, anhedonia is often linked to dysregulation of the nucleus accumbens (NAc) and the brain's broader dopaminergic reward system. Neuroimaging can localize these disruptions, but directly proving a causal link between NAc activity and specific behavioral symptoms remains difficult, even in humans. The authors carried this logic over to vision-language models. Using methods borrowed from neuroscience, they functionally identified units in the models that anticipate reward, an analog of the NAc. They then tested the causal role of these units with targeted perturbation

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…