Reward Valuation Vlm Anhedonia Causal

---
name: reward-valuation-vlm-anhedonia-causal
description: "Mechanistic framework identifying reward-anticipatory units in Vision-Language Models (VLMs) that parallel Nucleus Accumbens (NAc) function. Causal perturbation of NAc-selective units induces anhedonia-like behavioral shifts toward low-effort, low-reward options. Validates alignment between AI reward circuits and human dopaminergic reward system."
activation: "reward valuation, VLM anhedonia, Nucleus Accumbens, causal perturbation, reward anticipation, effort-based decision making, dopamine, motivational deficit, depression model"
tags: [neuroscience, vision-language-models, reward-system, anhedonia, causal-analysis, NAc, dopaminergic, depression]
version: 1.0.0
author: agent
arxiv_id: "2607.06626"
paper_title: "Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia"
---

# Reward Valuation in Vision-Language Models: Causal Mechanisms Underlying Anhedonia

## Core Innovation

### Problem
Recent Vision-Language Models (VLMs) capture increasingly complex aspects of human cognition. However, whether this alignment extends to **reward valuation** remains unclear. In the brain, anhedonia (loss of pleasure/motivation) is linked to dysregulation in the Nucleus Accumbens (NAc) and dopaminergic reward system.

**Challenge**: Neuroimaging has localized reward deficits, but establishing **causal links** between specific neural activity and behavioral symptoms remains difficult.

### Solution
Use neuroscience-inspired approach to:
1. **Functionally identify** reward-anticipatory units in VLMs
2. **Test causal role** via targeted perturbations
3. **Validate alignment** with human anhedonia mechanisms

## Methodology

### Step 1: Identify Reward-Anticipatory Units
```
Clinical Tests for Anhedonia
         ↓
Extract NAc activation patterns
         ↓
Search for analogous units in VLM
         ↓
Identify "NAc-selective units"
```

**Approach**:
- Use clinical tests developed to evaluate anhedonia in major depressive disorder
- Map NAc activity patterns to VLM internal representations
- Identify units that respond specifically to reward anticipation

### Step 2: Causal Perturbation
```
Baseline VLM Behavior
         ↓
Perturb NAc-selective units
         ↓
Observe behavioral changes
         ↓
Compare with human anhedonia symptoms
```

**Perturbation Types**:
- Ablation (zero activation)
- Amplification (increase activation)
- Suppression (decrease activation)

### Step 3: Behavioral Validation
Test whether perturbed VLM exhibits:
- Shift toward **low-effort, low-reward** options
- Preservation of task capability (not general impairment)
- Alignment with clinical anhedonia scales (DARS, MAP-SR)

## Key Findings

### 1. Causal Effect of NAc Perturbation
- Perturbing NAc-selective units induces behavioral effects **mirroring human anhedonia**
- Model shifts toward low-effort, low-reward options in effort-based decision-making tasks

### 2. Specificity of Deficit
- **Not a general capability loss**: Perturbed model maintains baseline performance when reward-based choice is removed
- **Specific to reward valuation**: Deficit targets reward anticipation, not task execution

### 3. Clinical Alignment
- Induced vulnerability aligns with clinical anhedonia scales:
  - **DARS** (Dimensional Anhedonia Rating Scale)
  - **MAP-SR** (Motivation and Pleasure Scale - Self Report)

### 4. Interpretation
Results reveal **reward valuation circuits in AI models** that parallel those in humans, suggesting:
- VLMs capture not just cognitive but motivational aspects of human cognition
- NAc function may be computationally principled enough to emerge in artificial systems
- Causal validation strengthens claims of brain-AI alignment

## Applications

### When to Use
- Studying reward processing in artificial systems
- Validating brain-AI alignment for motivational/affective circuits
- Developing computational models of anhedonia
- Testing interventions for reward-related psychiatric disorders
- Understanding how reward circuits emerge in large-scale models

### Implementation Steps
1. **Select clinical paradigm**: Choose validated anhedonia assessment (e.g., effort-based decision making)
2. **Identify neural signature**: Extract NAc activation pattern from human data or clinical literature
3. **Search VLM representations**: Use representational similarity analysis or probing classifiers
4. **Identify selective units**: Find units that respond to reward anticipation
5. **Design perturbation**: Choose ablation/amplification/suppression strategy
6. **Test behavioral effects**: Measure choices in effort-based tasks
7. **Validate specificity**: Confirm deficit is reward-specific, not general impairment
8. **Compare with clinical data**: Align with human anhedonia scales

### Pitfalls
- **Correlation vs. causation**: Must perturb to establish causal role
- **Specificity controls**: Must rule out general capability loss
- **Clinical validation**: Must compare with validated human measures
- **Unit selection**: May miss distributed representations
- **Task design**: Effort-based tasks must be sensitive to reward valuation

### Verification
- Test multiple perturbation types (ablation, amplification, suppression)
- Include control perturbations (non-NAc units)
- Validate on multiple effort-based decision tasks
- Compare with multiple clinical scales
- Test generalization across VLM architectures

## Biological Interpretation

### Nucleus Accumbens Function
- **Role**: Central to reward anticipation and motivation
- **Dysfunction**: Leads to anhedonia in depression, schizophrenia, addiction
- **Computational role**: Valuation of expected rewards vs. effort costs

### VLM Analogue
- **NAc-selective units**: Encode reward anticipation
- **Causal role**: Necessary for normal reward-based decision making
- **Alignment**: Suggests VLMs capture computationally principled aspects of reward processing

### Implications for Psychiatry
- **Computational psychiatry**: VLMs as testbeds for anhedonia mechanisms
- **Intervention design**: Target specific units/circuits for therapeutic effect
- **Biomarker development**: Identify neural signatures of reward dysfunction

## Limitations

### What This Does NOT Show
- VLMs do not "experience" anhedonia
- VLM reward circuits are not identical to biological NAc
- Findings do not directly translate to clinical interventions

### What This DOES Show
- VLMs capture computationally relevant aspects of reward processing
- Causal perturbation can reveal functional analogues of brain circuits
- Brain-AI alignment extends to motivational/affective systems

## References
- Honarmand, Aghabagher, Schrimpf (2026) "Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia" - arXiv:2607.06626
- Related: Nucleus Accumbens function in reward processing
- Anhedonia in major depressive disorder
- Brain-AI alignment in cognitive neuroscience