Skip to content

Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia

🕒 Published (v1): 2026-07-07 11:44 UTC · Source: Arxiv · link

Why this paper was selected

Schrimpf lab; mechanistic VLM–cognition alignment; novel anhedonia/reward valuation probe

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper investigates whether Vision-Language Models (VLMs) possess functional analogues to the brain's reward circuitry, specifically the Nucleus Accumbens (NAc). The authors identify NAc-selective units in VLMs via a neuroscience-inspired functional localization method, then demonstrate via activation patching that perturbing these units induces behavioral and psychometric profiles consistent with clinical anhedonia and motivational deficits seen in Major Depressive Disorder.

Problem

Brain-aligned AI research has focused primarily on sensory and early cognitive domains; whether AI models develop functional reward-anticipatory circuits—and whether their disruption produces human-interpretable psychiatric phenotypes—is unknown. Establishing a causal link between specific model units and reward-driven behavior would both deepen mechanistic interpretability and enable in silico simulation of affective brain disorders.

Method

The authors adapt the clinical Monetary Incentive Delay (MID) paradigm to VLMs: they extract MLP-block activations at the last token position across reward-incentivized vs. neutral prompts, then select units whose activation difference exceeds \(|\Delta| > 3\sigma\) as NAc-selective (top 0.7% of units in targeted layers, ~0.25% of the full model). Robustness is validated across four distinct linguistic framings of the reward cue (high cross-set cosine similarity). Layer analysis identifies two populations: late layers (18–27) with elevated reward-condition activation, and mid layers (13–14) with suppressed activation. The primary perturbation technique is activation patching: during reward trials, selected late-layer unit activations are replaced in-forward-pass with a pre-computed neutral-condition mean vector, leaving the rest of the model intact. The primary model is Qwen2-VL-7B-Instruct (28-layer decoder-only transformer with integrated vision encoder). Behavioral evaluation uses two adapted EEfRT variants (ASDiv-EEfRT with math-problem difficulty tiers, and Probability-EEfRT mirroring the clinical risk paradigm) and three clinical psychometric scales (DARS, MAP-SR, AES).

Key Contributions

  • First causal identification and perturbation of reward-anticipatory circuits in VLMs, using neuroscience methodology (MID paradigm adaptation, \(3\sigma\) threshold functional localization).
  • Demonstration of a specific anhedonic behavioral phenotype—shift toward low-effort/low-reward choices—without general cognitive degradation (preserved forced-choice accuracy).
  • Psychometric alignment of the perturbed model with MDD clinical profiles across DARS, MAP-SR, and AES scales.
  • Parametric dose-response validation: monotonic motivation collapse as NAc-unit scaling factor \(\lambda\) decreases.
  • Perspective-shifting dissociation: the perturbed model is impaired when choosing for itself but remains intact when choosing for another agent, separating reward valuation from conceptual reward understanding.
  • Novel computational EEfRT variants (ASDiv-EEfRT, Probability-EEfRT) enabling objective behavioral measurement of incentive motivation in LMs.

Results

  • DARS: NAc-perturbed model scores drop 16.7% vs. intact (\(p \ll 0.001\)), reflecting shift to diminished anticipatory interest.
  • MAP-SR: 2.4% score reduction (\(p \ll 0.001\)) in frequency/intensity of expected pleasure.
  • AES (inverted): 8.6% drop (\(p \ll 0.001\)), indicating significant increase in apathy.
  • ASDiv-EEfRT: perturbed model significantly shifts toward low-point (easy) math problems; mean points chosen are significantly reduced vs. intact model.
  • Probability-EEfRT: significant reduction in high-effort task selection in NAc-perturbed vs. intact model.
  • Control—random perturbation: perturbing an equivalent number of randomly selected units produces no significant change in choice frequency, confirming specificity of NAc-unit effect.
  • Control—reasoning capability: no significant difference in forced-choice (no-reward) accuracy between intact and perturbed models, confirming preserved general cognition.
  • Parametric scaling: monotonic dose-response collapse in motivation as scaling factor decreases to minimum; mid-layer patching does not induce anhedonia.
  • Score reductions directionally mirror human MDD vs. healthy control differences on all three scales (though human group differences lack statistical significance due to high variance).

Limitations

  • Primary model is a single architecture (Qwen2-VL-7B-Instruct); results for other models are relegated to the appendix.
  • MID task adaptation uses purely textual stimuli despite the model being a VLM; visual reward cues are not tested.
  • Self-report scales (DARS, MAP-SR, AES) are inherently adapted for language models and may not faithfully reproduce the phenomenological dimension of clinical instruments.
  • MAP-SR score shift (2.4%) is notably smaller than the magnitude seen in human MDD populations.
  • Human clinical group comparisons do not reach statistical significance due to high inter-subject variability, limiting direct quantitative benchmarking against human data.
  • Causality is established within the forward pass via patching; upstream circuit interactions and compensatory dynamics over continued inference are not analyzed.

Relevance to Vision-Language Models

This work directly advances mechanistic interpretability of VLMs by showing that multimodal task optimization gives rise to latent circuits for reward valuation, not just perception or language. For researchers tracking VLMs, it introduces a functional-localization methodology—adapted from clinical fMRI paradigms—that is broadly applicable to probing other cognitive/affective constructs in large multimodal models. The finding that a small fraction (~0.25%) of units causally mediates reward-driven decision-making has implications for understanding emergent behavior in VLMs beyond their training objectives. It also opens a line of inquiry into whether VLMs' internal representations encode motivationally relevant states that could influence reliability, goal-directedness, or alignment properties in agentic deployments.