Reward Valuation in Large Language Models: Causal Induction of Anhedonia
Researchers found that specific reward-anticipatory units within vision-language models, analogous to human brain circuits, can be perturbed to induce behavioral effects mirroring anhedonia. This manipulation causes the AI to shift toward low-effort options when rewards are present but maintains general task capability otherwise. This establishes a causal link between internal reward representations and motivational deficits.