Learnijoy
ResearchAI ResearchAI Engineering & DevTools

AI Models Show Anhedonia-Like Behavior Through Reward Valuation Circuits.

Melika Honarmand, Samin Mahdipour Aghabagher, Martin Schrimpf· July 9, 2026 View original

Summary

Researchers identified reward-anticipatory units in Vision-Language Models (VLMs) that, when perturbed, induce behaviors mirroring human anhedonia, a deficit in experiencing pleasure. This suggests VLMs possess internal reward valuation circuits analogous to those in the human brain.

This research explores whether Vision-Language Models (VLMs) exhibit reward valuation mechanisms similar to human cognition, specifically focusing on anhedonia, a symptom of major depressive disorder characterized by a reduced ability to experience pleasure. Drawing inspiration from neuroscience, where anhedonia is linked to dysregulation in the Nucleus Accumbens (NAc) and dopaminergic reward systems, the study sought to find analogous structures in AI. The team functionally identified specific "reward-anticipatory units" within VLMs. When these NAc-selective units were targeted and perturbed, the models displayed behavioral changes consistent with human anhedonia. Specifically, the perturbed models favored low-effort, low-reward options in decision-making tasks, while maintaining baseline performance when reward-based choices were removed. This indicates a specific deficit in reward valuation and anticipation, rather than a general loss of task capability, suggesting that AI models can develop internal circuits for reward processing that parallel human neurological functions.

Why it matters

This research offers a novel way to understand and potentially diagnose complex cognitive states like anhedonia in AI, opening doors for more human-aligned AI development and potentially new insights into human neuroscience.

How to implement this in your domain

  1. 1Develop AI systems with explicit, controllable reward valuation modules to prevent unintended anhedonia-like behaviors in critical applications.
  2. 2Design AI agents that can adapt their reward functions based on observed "motivational" states to improve long-term performance and robustness.
  3. 3Utilize mechanistic interpretability techniques to map and understand reward pathways in complex AI models for better control and ethical deployment.
  4. 4Explore the implications of these findings for AI safety, ensuring AI systems maintain appropriate motivation and goal-seeking behaviors.

Who benefits

AI EthicsHealthcare (AI for mental health research)RoboticsGame Development

Key takeaways

  • Vision-Language Models exhibit internal reward valuation circuits.
  • Perturbing specific VLM units can induce anhedonia-like behaviors.
  • The observed deficit is in reward valuation, not general task capability.
  • This research bridges AI interpretability with neuroscience insights into human cognition.

Original post by Melika Honarmand, Samin Mahdipour Aghabagher, Martin Schrimpf

"arXiv:2607.06626v1 Announce Type: new Abstract: Recent Vision-Language Models capture increasingly complex aspects of human cognition. Here we ask whether this alignment extends to reward valuation, which we assess in a mechanistic framework built on clinical tests that were deve…"

View on X

Originally posted by Melika Honarmand, Samin Mahdipour Aghabagher, Martin Schrimpf on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

Transformers Learn Non-Invertible Modular Multiplication via Stratified Fourier Mechanisms.

This research investigates how small transformers learn modular integer multiplication over composite moduli, a non-invertible operation. It proposes the "monoid extension" theory, suggesting models partition input space into hierarchical algebraic regions where Fourier mechanisms apply, explaining how embeddings, attention, and local features contribute to the computation.

Zitong Andrew Chen, Junaid Hasan, Akhil Srinivasan, Hemkesh Bandi, Jarod AlperJul 9, 2026
AI Engineering & DevToolsAI Research

New Interpretable Model Handles Feature Interactions in Tabular Data.

This paper introduces Interaction Aware Interpretable Machine Learning (IAIML), a framework for tabular data that addresses the limitation of traditional interpretable models in capturing feature interactions. IAIML uses adaptive discretization, pairwise interaction scoring, and a partitioned explanation budget to achieve high accuracy while maintaining interpretability.

Srikumar KrishnamoorthyJul 9, 2026
AI ResearchAI Engineering & DevTools

Principles of Deep Feedforward ReLU Networks Unveiled.

This paper systematically studies the mechanisms of deep feedforward ReLU networks, generalizing principles from two-layer networks to deeper architectures. It explains how hidden-layer units form piecewise linear manifolds to divide input space and how paths and their relationships are central to understanding the back-propagation training solution.

Changcun HuangJul 9, 2026