SciGroveBeta
Neuroscience

Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia

Melika Honarmand, Samin Mahdipour Aghabagher, Martin Schrimpf

Featured July 14, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

AI models that understand both pictures and words can develop internal 'reward centers' that, when disrupted, make them lose their motivation to work hard for big rewards, just like people with anhedonia.

In depth
The paper demonstrates that Vision-Language Models (VLMs) develop internal reward-anticipatory units that functionally resemble the human Nucleus Accumbens (NAc). By applying targeted activation patching to these units, the authors causally induce anhedonia-like behavior in the models, characterized by a significant shift towards low-effort, low-reward choices, while preserving general cognitive capabilities.

Key Takeaways

  • 1
    Vision-Language Models develop reward valuation circuits that functionally align with the human brain's Nucleus Accumbens.
  • 2
    Targeted perturbation of these units causally induces anhedonia-like behavior in VLMs, leading to a preference for low-effort, low-reward options.
  • 3
    The induced anhedonia represents a selective motivational deficit, as the perturbed models maintain their reasoning abilities in non-reward contexts.

Conceptual Flow

HIGH LEVEL
1
Methodology: How was it done?

The researchers found special parts in an AI model that light up when it expects a reward, just like a brain's reward center. Then, they gently 'turned down' these parts to see what happened.

AI Model
Reward Task
Find Reward Parts
Turn Down Reward Parts
2
Results: What did they find?

After turning down the reward parts, the AI model stopped wanting to do hard tasks for big rewards, preferring easy tasks, showing it lost its 'get up and go.'

AI Model (Reward Parts Down)
Hard Task Option
Easy Task Option
Choose Task
Always Picks Easy Task