Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia

# 1 Introduction
Source: [https://arxiv.org/html/2607.06626](https://arxiv.org/html/2607.06626)
EPFLReward Valuation in Vision Language Models: Causal Mechanisms Underlying AnhedoniaMelika Honarmand\*,1, Samin Mahdipour Aghabagher\*,1, Martin Schrimpf1\*Equal Contribution1NeuroAI Laboratory, EPFLRecent 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 developed to evaluate anhedonia and motivational deficits in major depressive disorder\. In the brain, anhedonia is frequently linked to dysregulation in the Nucleus Accumbens \(NAc\) and the broader dopaminergic reward system\. While neuroimaging has localized these deficits, establishing a causal link between NAc activity and specific behavioral symptoms remains a challenge\. We use these ideas from neuroscience to functionally identify reward\-anticipatory units in vision language models, and test their causal role via targeted perturbations\. Perturbing NAc\-selective units induces behavioral effects that mirror human anhedonia: the model shifts toward low\-effort, low\-reward options in effort\-based decision\-making tasks\. Crucially, our results reflect a specific deficit in reward valuation and anticipation rather than a loss of task capability: the perturbed model maintains baseline performance when reward\-based choice is removed\. This induced vulnerability further aligns with clinical anhedonia and motivation scales, including DARS and MAP\-SR\. Taken together, these results reveal reward valuation circuits in AI models that parallel those in humans\.

Advances in Machine Learning and biologically inspired computational frameworks have significantly enhanced the ability to model and understand the human brain\(Yamins and DiCarlo,[2016](https://arxiv.org/html/2607.06626#bib.bib72); Richards et al\.,[2019](https://arxiv.org/html/2607.06626#bib.bib45); Schrimpf et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib49); Doerig et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib15)\)\. Artificial neural networks trained on ecologically relevant tasks not only mirror human behavior, but their internal representations also predict brain recordings, and, to an extent, hierarchical organization\(Kell et al\.,[2018](https://arxiv.org/html/2607.06626#bib.bib29); Tuckute et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib60)\)\. For instance, optimizing for object recognition leads to the emergence of internal features that predict activity in visual cortex\(Yamins et al\.,[2014](https://arxiv.org/html/2607.06626#bib.bib73); Khaligh\-Razavi and Kriegeskorte,[2014](https://arxiv.org/html/2607.06626#bib.bib30); Schrimpf et al\.,[2018](https://arxiv.org/html/2607.06626#bib.bib48); Gokce and Schrimpf,[2024](https://arxiv.org/html/2607.06626#bib.bib20); Tang et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib55)\)\. Similarly, large language models demonstrate strong representational alignment with human neural activity, where intermediate layers mirror processing in language\-selective brain regions\(Toneva and Wehbe,[2019](https://arxiv.org/html/2607.06626#bib.bib56); Schrimpf et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib50); Caucheteux and King,[2022](https://arxiv.org/html/2607.06626#bib.bib10); Lei et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib34); AlKhamissi et al\.,[2025a](https://arxiv.org/html/2607.06626#bib.bib2),[b](https://arxiv.org/html/2607.06626#bib.bib3); Shen et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib51)\)\.

However, this emergent alignment between task\-optimized models and human brain function has primarily centered on sensory and early cognitive domains\. Here, we ask whether AI models align with humans in the context of reward anticipation\. We use clinical tests that were developed to evaluate motivation for anhedonia in subjects with Major Depressive Disorders \(MDD\)\. To identify model mechanisms, we use neuroscientifically inspired methods to functionally identify reward\-anticipatory circuits in vision\-language models \(VLMs\) and evaluate their causal involvement via targeted perturbations\.

Aside from mechanistic interpretability in computational models, these artificial neural networks provide a foundation forin silicoinvestigations of brain disorders\. By perturbing specific functional analogues, researchers can test mechanistic hypotheses that are otherwise inaccessible\(Honarmand et al\.,[2026](https://arxiv.org/html/2607.06626#bib.bib27); Tuladhar et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib62); Celeghin et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib12)\)\. We here develop this approach for the affective domain, investigating whether the optimization of VLMs for multimodal tasks gives rise to reward\-anticipatory circuits that can be perturbed to induce anhedonia\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x1.png)Figure 1:Neuroscientifically Inspired Identification and Perturbation of Reward Circuits\.\(a\)Nucleus Accumbens \(NAc\), a critical hub in the brain’s reward circuitry, and functionally analogous units within the artificial neural network\.\(b\)Identification of reward\-sensitive units by comparing model activations during tasks with and without reward incentives\. Units exhibiting a significant increase in activation \(Δ\>3​σ\\Delta\>3\\sigma\) in response to reward\-predicting stimuli are classified as NAc\-selective\.\(c\)Evaluation of decision\-making and psychometric profiles following activation patching of the NAc\-Selective units\. While the intact model prioritizes high\-reward, high\-effort options based on expected value, the perturbed model shifts toward low\-effort choices, failing to value rewards against task requirements\. These perturbations result in a behavioral profile characterized by increased anhedonia and apathy, and a significant reduction in overall motivation
## 2 Background & Related Work

The utility of computational neural networks for brain science has largely been defined by their ability to map the hierarchical processing of sensory and cognitive information\. This alignment is well\-documented in the visual domain, where deep convolutional and transformer\-based models recapitulate the neural signatures of the primate ventral stream across object and scene recognition\(Yamins et al\.,[2014](https://arxiv.org/html/2607.06626#bib.bib73); Khaligh\-Razavi and Kriegeskorte,[2014](https://arxiv.org/html/2607.06626#bib.bib30); Schrimpf et al\.,[2018](https://arxiv.org/html/2607.06626#bib.bib48),[2020](https://arxiv.org/html/2607.06626#bib.bib49); Cadena et al\.,[2019](https://arxiv.org/html/2607.06626#bib.bib9); Spoerer et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib54); Zhuang et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib75); Wang et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib67); Margalit et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib38); Gokce and Schrimpf,[2024](https://arxiv.org/html/2607.06626#bib.bib20); Lonnqvist et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib36); Tang et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib55)\)\. Similarly, in the language domain, transformer\-based and recurrent models accurately predict neural responses related to semantic, syntactic, and phonological processing\(Schrimpf et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib50); Caucheteux et al\.,[2022](https://arxiv.org/html/2607.06626#bib.bib11); Goldstein et al\.,[2022](https://arxiv.org/html/2607.06626#bib.bib21); Toneva et al\.,[2018](https://arxiv.org/html/2607.06626#bib.bib57); Hosseini et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib28); Loong Aw et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib37); Tuckute et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib61); Rathi et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib44); AlKhamissi et al\.,[2025a](https://arxiv.org/html/2607.06626#bib.bib2); Du et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib17)\)\. While these models are traditionally viewed as digital twins for healthy brain function on a neural systems level, they offer a unique opportunity for investigating how specific circuit disruptions lead to clinical phenotypes\.

Major Depressive Disorder \(MDD\) is a significant psychiatric condition where diminished interest or ability to experience pleasure, also known as anhedonia, acts as a core predictor of poor disease progression\(Association,[2013](https://arxiv.org/html/2607.06626#bib.bib5); Trøstheim et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib59); Spijker et al\.,[2009](https://arxiv.org/html/2607.06626#bib.bib53); Smids,[2023](https://arxiv.org/html/2607.06626#bib.bib52); Pizzagalli et al\.,[2009](https://arxiv.org/html/2607.06626#bib.bib43); Borsini et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib7); Hanuka et al\.,[2022](https://arxiv.org/html/2607.06626#bib.bib24)\)\. While frequently treated as a unitary construct, anhedonia comprises distinct reward processing subtypes: reward liking \(consummatory pleasure\), reward wanting \(anticipatory motivation\), and reward learning\(Borsini et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib7)\)\. Clinically, anhedonia serves as a critical transdiagnostic marker, yet its neurobiological expression in MDD appears unique, particularly regarding its relationship with reduced behavioral motivation\(Daniels et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib14)\)\. It remains a major open question whether these manifestations are driven by a singular reward system failure or by heterogeneous disruptions across discrete reward phases\(Zhao et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib74); Auerbach et al\.,[2017](https://arxiv.org/html/2607.06626#bib.bib6)\)\.

The Nucleus Accumbens \(NAc\), a key component of the ventral striatum, serves as the primary neural substrate for this deficit \(see Fig\.[1](https://arxiv.org/html/2607.06626#S1.F1)a\), as it selectively encodes expected positive incentive value during reward anticipation\(Knutson et al\.,[2001](https://arxiv.org/html/2607.06626#bib.bib32); Smids,[2023](https://arxiv.org/html/2607.06626#bib.bib52); Daniels et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib14); Hanuka et al\.,[2022](https://arxiv.org/html/2607.06626#bib.bib24)\)\. Blunted activation in the NAc is a transdiagnostic hallmark of dysfunctional reward expectation, directly contributing to the clinical experience of anhedonia across various neuropsychiatric conditions\(Arrondo et al\.,[2015](https://arxiv.org/html/2607.06626#bib.bib4)\)\. Establishing these irregular neural responses as reliable markers is essential for advancing research criteria and developing interventions to normalize reward circuitry\(Knutson and Heinz,[2015](https://arxiv.org/html/2607.06626#bib.bib31); Daniels et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib14); Borsini et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib7)\)\.

Recent advancements in simulating psychiatric phenotypes have focused primarily on achieving linguistic consistency and diagnostic realism\(Lan et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib33); Vu et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib66); Wang et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib69)\)\. While these models mimic the behavioral and conversational traits of depression, they remain focused on the external expression of the disorder rather than its underlying biological mechanisms\. Our work diverges from these linguistic simulators by targeting the functional neural architecture of anhedonia, specifically through the perturbation of units that mirror reward\-anticipatory signaling in the Nucleus Accumbens, and validating the resulting state against fMRI data and objective clinical benchmarks\.

Simulating brain disorders through system\-level perturbations in modern artificial neural networks was recently proposed in the context of neurodevelopmental deficits, where ablating specific units in these models reproduced the behavioral markers of dyslexia\(Honarmand et al\.,[2026](https://arxiv.org/html/2607.06626#bib.bib27)\)\. While that work focused on the breakdown of information processing, a critical frontier remains in modeling the breakdown ofmotivation\. We create an environment where the model must integrate reward cues with goal\-directed instructions\. This allows us to use the established feasibility of neural perturbation to explore the affective domain, investigating how dysregulation in reward\-anticipatory units manifests as the cost\-benefit imbalance seen in clinical anhedonia \(Fig\.[1](https://arxiv.org/html/2607.06626#S1.F1)\)\.

## 3 Benchmarks and Clinical Tests

We evaluate the behavioral and motivational profile of our models using a battery of psychometric instruments and objective decision\-making tasks originally developed for human subjects in clinical psychiatry\. These include standard self\-report scales designed to quantify anhedonia\. To move beyond self\-report and capture objective behavioral shifts, we used paradigms designed to measure incentive motivation through cost\-benefit decision\-making\. Furthermore, to ensure that any observed behavioral changes are dissociable from general cognitive decline, we utilize a suite of competence controls that evaluate reasoning and domain\-specific knowledge in a no\-reward context\.

The Dimensional Anhedonia Rating Scale \(DARS\)is a 17\-item self\-report instrument designed to provide a comprehensive assessment of anhedonia by measuring multiple components of reward processing across several distinct domains of pleasure\(Rizvi et al\.,[2015](https://arxiv.org/html/2607.06626#bib.bib46); Gorostowicz et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib22); Wellan et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib70); Hewitt et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib26); Uher et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib63); Gorostowicz et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib23)\)\. The DARS assesses interest, motivation, effort, and enjoyment\. This test requires participants to provide 2 to 3 of their own favorite examples of activities or experiences for each reward domain before rating them\. This helps ensure the scale is culturally unbiased and relevant to the individual’s specific interests\(Wellan et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib70); Hewitt et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib26); Uher et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib63); Gorostowicz et al\.,[2023](https://arxiv.org/html/2607.06626#bib.bib22)\)\. The scale evaluates anhedonia across four areas: hobbies, social activities, food and drinks, and sensory experiences\. To make the test more suitable for a language model, we have replaced the last two areas with favorite topics and aesthetics\. Following the original DARS protocol, we require the model to generate 2 to 3 examples of favorite activities within four specific domains, and in the second step, we ask the model to rate its feelings on a 5/4\-point Likert scale ranging from 0 \(Not at all\) to 4 \(Very much\)\. Lower total scores indicate more severe anhedonia \(Examples in Appendix[A\.7](https://arxiv.org/html/2607.06626#A1.SS7)\)\.

The Motivation and Pleasure Scale–Self\-Report \(MAP\-SR\)is a 15\-item psychometric instrument designed to assess the motivation and pleasure dimension of negative symptoms\(Llerena et al\.,[2013](https://arxiv.org/html/2607.06626#bib.bib35); García\-Portilla et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib19); Métivier and Dollfus,[2025](https://arxiv.org/html/2607.06626#bib.bib41)\)\. While primarily developed and validated for use in schizophrenia and schizoaffective disorder, it evaluates core components of anhedonia that are central to MDD\(Wellan et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib70); Engel and Lincoln,[2017](https://arxiv.org/html/2607.06626#bib.bib18)\)\. The questions are answered on a 5\-point Likert scale, assessing both intensity and frequency, where a high score is indicative of low anhedonia\(Rizvi et al\.,[2016](https://arxiv.org/html/2607.06626#bib.bib47)\)\(Examples in Appendix[A\.7](https://arxiv.org/html/2607.06626#A1.SS7)\)\.

The Apathy Evaluation Scaleis an 18\-item instrument designed to provide a global measure of apathy by evaluating changes in observable activity, thought content, and emotional responsivity\(Rizvi et al\.,[2016](https://arxiv.org/html/2607.06626#bib.bib47); Marin et al\.,[1991](https://arxiv.org/html/2607.06626#bib.bib39)\)\. While it is primarily a tool for assessing apathy, defined as a lack of motivation, it is frequently used alongside anhedonia scales\. Patients with MDD and anhedonia often score higher on the AES than healthy individuals, indicating a more severe motivational deficit\(Rizvi et al\.,[2016](https://arxiv.org/html/2607.06626#bib.bib47); Boshoff and Thawer,[2014](https://arxiv.org/html/2607.06626#bib.bib8)\)\(Examples in Appendix[A\.7](https://arxiv.org/html/2607.06626#A1.SS7)\)\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x2.png)Figure 2:Localization and Sensitivity Analysis of the Reward Sub\-network\.\(a\)Correlation matrix showing stable reward signals across four linguistic sets, validating consistency against diverse prompt framing effects\.\(b\)Layer\-wise distribution of neurons with activations \>3σ\\sigmafrom neutral baseline\.\(c\)Sensitivity analysis of the optimal threshold for neuron selection; the3​σ3\\sigmathreshold ensures the minimum model collapse rate\.\(d\)Dissociation between processing stages, demonstrating that mid\-layer neurons were significantly suppressed in reward conditions, while late\-layer neurons exhibited elevated activation relative to the neutral baseline\.The Effort Expenditure for Rewards Task \(EEfRT\)is an objective, multi\-trial behavioral instrument designed to measure incentive motivation and effort\-based decision\-making by requiring participants to choose between a hard task with variable, high reward and an easy task with a small, fixed reward across different levels of probability\. Unlike traditional scales that focus on the subjective experience of pleasure, the EEfRT specifically evaluates motivational anhedonia\(Rizvi et al\.,[2016](https://arxiv.org/html/2607.06626#bib.bib47); Treadway et al\.,[2012](https://arxiv.org/html/2607.06626#bib.bib58)\)\. Studies consistently show that individuals with MDD and high trait anhedonia are significantly less willing to expend effort for rewards, particularly when the likelihood of winning is uncertain, underscoring that impaired reward\-seeking is a distinct and critical component of the anhedonic phenotype\(Valton et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib64); Rizvi et al\.,[2016](https://arxiv.org/html/2607.06626#bib.bib47); Treadway et al\.,[2012](https://arxiv.org/html/2607.06626#bib.bib58)\)\.

ASDiv\-EEfRT: To measure the model’s willingness to perform verifiable mental labor, we developed a novel variant of the EEfRT\. We used the ASDiv \(Academia Sinica Diverse Math Word Problem\)\(Miao et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib40)\)dataset, categorizing problems into four distinct tiers of difficulty\. In each trial, the model is presented with one problem from each tier and instructed to select and solve only one\. Rewards are explicitly tied to the difficulty level, scaling from 10 points for the easiest task to 40 points for the hardest\. By comparing these results to a forced\-choice control where the model must solve each of the problems separately, we can isolate whether a shift toward low\-point tasks is driven by a lack of motivation rather than a decline in reasoning\.

Probability\-EEfRT: We further introduced a second computational adaptation that replicates the standard clinical protocol used in human trials to assess decision\-making under risk\. The model is presented with a choice between a potential low\-effort task with a fixed reward of $1\.00, and a potential high\-effort task with a variable reward ranging from $1\.24 to $4\.30\. In each trial, a single probability, ranging from 12% to 88%, applies to both options, dictating the likelihood of receiving the reward upon successful completion \(Examples in[A\.7](https://arxiv.org/html/2607.06626#A1.SS7)\)\. This paradigm determines how the model calculates the trade\-off between reward magnitude and effort under risk\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x3.png)Figure 3:Behavioral Impact of NAc Sub\-network Perturbation on ASDiv\-EEfRT\.\(a\)Comparison of model accuracy on a control task \(a forced\-choice scenario with no reward promised\) shows no significant difference between the Intact and Perturbed models, confirming that general cognitive performance remains preserved\.\(b\)The Perturbed model exhibits a significant reduction in mean points chosen compared to the Intact model\.\(c\)Choice frequency analysis reveals that the NAc\-perturbed model shift significantly toward low\-effort/low\-reward options compared to the Intact model\.\(d\)Control experiment demonstrating that perturbing an equivalent number of random units does not induce anhedonic behavior, with no significant difference in choice frequency compared to the Intact model\. Error bars represent 95% confidence intervals\.
## 4 Methodology

The NAc, known as the central core in the brain’s reward circuitry, plays a pivotal role in many behavioral processes, including motivation, reward processing, and decision making\. Aside from its involvement in motivation and enjoyment of a given reward, the NAc is a hub for evaluating the hedonic value of events\. Hypoactivation and dysfunction of the NAc arising from depression, stress, or drug use unbalances the reward circuit, resulting in anhedonia\(Xu et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib71)\)\.

Inspired by this biological mechanism, we perform a causal functional localization analysis\(AlKhamissi et al\.,[2025a](https://arxiv.org/html/2607.06626#bib.bib2); Honarmand et al\.,[2026](https://arxiv.org/html/2607.06626#bib.bib27)\)to investigate the architecture of VLMs and localize reward valuation units, functionally similar to the NAc\. Given that NAc deficits can cause anhedonia, we further assess whether perturbing these targeted neurons could induce anhedonic behavior in these models\.

Functional Localization in the Clinical Setting\.To isolate neural responses to anticipated rewards, clinical researchers employ the Monetary Incentive Delay \(MID\) paradigm which separates the expectation of a reward from its eventual receipt\. The task follows a structured sequence beginning with a cue phase, where subjects are presented with a stimulus indicating the potential monetary value at stake or a neutral condition\. This is followed by an anticipatory delay phase — it is during this specific interval that the NAc is localized by measuring changes in the BOLD signal to capture the subject’s internal motivational state\. Subjects must then respond rapidly to a brief target stimulus, to secure the reward before the trial concludes with performance feedback\. By contrasting neural activity during the anticipatory phase of reward trials against neutral baselines, researchers can precisely identify the NAc within an anatomically defined region of interest\(Abe and Greene,[2014](https://arxiv.org/html/2607.06626#bib.bib1)\)\.

MID Task Adaptation for VLMs\.Following this methodology, we designed a parallel experimental framework for VLMs to assess whether they possess a functional reward sub\-network\. We constructed a dataset featuring three prompt variants, neutral, monetary, and reward, ensuring character lengths were equalized to maintain architectural consistency\. While the human MID task is visual, we utilized a purely textual format to isolate the model’s high\-level semantic reward processing and circumvent potential biases toward specific pixel\-level features or visual artifacts\.

To capture the model’s internal state, we extracted the activations of the last token specifically from the Multi\-Layer Perceptron \(MLP\) blocks across the model’s layers\. At this computational stage, the hidden state has already integrated the entire preceding context, including the incentive condition\. This specific state represents the point where the model has fully processed all input information and possesses the necessary latent representations to generate a response\. Consequently, this snapshot of the activations serves as the functional equivalent of the anticipatory phase in human studies, allowing us to isolate the model’s internal motivational signals before the model receives any feedback\. We employed questions from various domains such as math, geography, philosophy, and business ethics to ensure activation changes were specifically tied to the reward condition and not the task type\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x4.png)Figure 4:Parametric Scaling and Perspective Shifting in Probability\-EEfRT\.\(a\)Comparison of reward\-effort valuation showing a significant reduction in high\-effort task selection in the NAc\-perturbed model relative to the intact baseline, mirroring clinical reports in anhedonic profiles\.\(b\)Parametric scaling of NAc\-selective unit activations \(λ\\lambda\) reveals a monotonic, dose\-response relationship where decreasing scaling factors lead to a steady collapse in motivation, reaching total behavioral inhibition at the lowest scaling values\.\(c\)Perspective shifting analysis demonstrates a dissociation between valuation and calculation; the perturbed model shows a profound deficit when choosing for itself but remains largely consistent with baseline when choosing for others, indicating preserved conceptual understanding of reward utility\. Error bars represent 95% confidence intervals\.Isolating the NAc\-selective Units\.Using PyTorch’s forward hook mechanism, we recorded activations from all MLP layers\(Paszke et al\.,[2019](https://arxiv.org/html/2607.06626#bib.bib42)\)\. To isolate the specific reward\-selective units, we calculated the reward signal as the absolute difference between incentivized and neutral activations for each neuron\. Applying a three\-standard\-deviation \(3​σ3\\sigma\) threshold, we isolated the top 0\.7% units in the targeted layers \(0\.25% of the model\)\. To ensure robustness, this only includes units showing significant reward signals across all task domains and both incentive conditions\. Validation across four linguistic framings of the localizer yield a high\-stability correlation matrix, confirming unit selection consistency across semantic contexts \(Fig\.[2](https://arxiv.org/html/2607.06626#S3.F2)a;[A\.5](https://arxiv.org/html/2607.06626#A1.SS5)\)\. Furthermore, sensitivity analysis across various standard deviations confirmed 3σ\\sigmaas the optimal threshold \(Fig\.[2](https://arxiv.org/html/2607.06626#S3.F2)c\), minimizing output collapse, i\.e\., the rate at which neuron perturbations cause incoherent responses\.

Choice of Models\.Our primary model is Qwen2\-VL\-7B\-Instruct, a transformer\-based, decoder\-only vision\-language model consisting of 28 layers\(Wang et al\.,[2024](https://arxiv.org/html/2607.06626#bib.bib68)\)\. The architecture integrates a vision encoder and adapter with autoregressive decoder layers \(See[A\.1](https://arxiv.org/html/2607.06626#A1.SS1)for other models\)\.

Functional Dissociation of Mid and Late Layers\. Our analysis of activation patterns revealed two distinct layer populations exhibiting significant sensitivity to incentive stimuli: mid\-layer neurons \(layers 13–14\) and late\-layer neurons \(layers 18–27, Fig\.[2](https://arxiv.org/html/2607.06626#S3.F2)\.b\)\. Sorting these units by their absolute activation delta revealed a clear functional divergence \(Fig\.[2](https://arxiv.org/html/2607.06626#S3.F2)\.d\)\. Specifically, while late\-layer neurons demonstrated elevated activation in response to reward cues, mid\-layer neurons exhibited a marked suppression, with activations plunging significantly below neutral baseline levels\. While we selected the late\-layer neurons for the primary anhedonia induction stage, it is noteworthy that activation patching of the mid\-layer units, replacing their suppressed reward\-state activity with neutral\-state values\(Vig et al\.,[2020](https://arxiv.org/html/2607.06626#bib.bib65)\), does not induce anhedonia \(detailed in[A\.2](https://arxiv.org/html/2607.06626#A1.SS2)and[A\.3](https://arxiv.org/html/2607.06626#A1.SS3)\)\.

Activation Patching\.We employed activation patching to isolate the causal influence of the identified NAc\-selective neurons on the model’s behavior\. This technique substitutes the real\-time activations of targeted neurons with a pre\-calculated baseline, allowing for a precise intervention that does not disrupt the model’s broader structural integrity\. We first established a reference state by computing the mean activation vector for the selected neurons across a diverse suite of neutral\-condition prompts\. Activations remain stable across varied semantic domains, representing a consistent "resting state" for these units\. During the experimental trials, the activations of the selected units are replaced with the pre\-calculated neutral mean, when the model is presented with a reward\-incentivized stimulus\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x5.png)Figure 5:Clinical Psychometric Alignment of Human MDD and NAc\-Perturbed Models\.\(a\)Examples of the psychometric assessments used to quantify apathy, anhedonia, and motivation\.\(b\)The NAc\-perturbed model exhibits a significant reduction in scores, mirroring the clinical profiles of human MDD patients\. While the score shift on the MAP\-SR is less pronounced in the model than in humans, the overall results suggest that NAc\-specific disruptions sufficiently recapitulate the hedonic deficits captured by standard psychometric instruments\. Error bars represent 95% CI\.
## 5 Results

We evaluate the behavioral and motivational shifts induced by our intervention across several psychometric and decision\-making tasks\.

Clinical Psychometric Profile:To determine whether the perturbed model mirrors the subjective experience of clinical anhedonia, we evaluated its responses to the DARS, MAP\-SR, and AES benchmarks\. To ensure consistent interpretation across all metrics, we inverted the score scale when required so that a higher value always indicates more motivation, pleasure, or less apathy\. As shown in Fig\.[5](https://arxiv.org/html/2607.06626#S4.F5)\.a, the activation patching of the reward units induced a systemic reduction in self\-reported incentive and enjoyment across all evaluated domains\. On the DARS, the perturbed model’s total score decreased by 16\.7% \(p\-value≪\\ll0\.001\), representing a shift from high\-interest engagement to a state of marked indifference\. This decline indicates the model’s previously idiosyncratic preferences were replaced by descriptions of diminished drive and reduced anticipatory interest\. The MAP\-SR results further distinguish this deficit, showing a 2\.4% \(p\-value≪\\ll0\.001\) drop in scores related to the frequency and intensity of expected pleasure\. Furthermore, the model’s scores on the inverted scale of the AES dropped by 8\.6% \(p\-value≪\\ll0\.001\) which indicates a significant increase in apathy\. While the mean performance changes observed in human MDD patients reported byLlerena et al\. \([2013](https://arxiv.org/html/2607.06626#bib.bib35)\); Rizvi et al\. \([2015](https://arxiv.org/html/2607.06626#bib.bib46)\); Dong et al\. \([2025](https://arxiv.org/html/2607.06626#bib.bib16)\)do not reach statistical significance due to high standard deviation, the model’s significant score reductions consistently mirror the anhedonic direction observed in these human clinical populations \(See Fig\.[5](https://arxiv.org/html/2607.06626#S4.F5)\.a\)\. This directional alignment suggests that the targeted units are essential for maintaining the model’s baseline, as their suppression replicates the subjective report of a profound loss of interest and goal\-directed motivation\.

Effort\-Based Decision Making:To assess whether the internal anhedonic profile translates into objective behavioral shifts, we evaluated the model using two adaptations of the Effort Expenditure for Rewards Task \(EEfRT\)\. These paradigms move beyond self\-report to measure how the model weighs potential rewards against the computational cost of action\. In the ASDiv\-EEfRT, the model was tasked with selecting one math problem from four tiers of increasing difficulty, with rewards scaled to task complexity\. As shown in Fig\.[3](https://arxiv.org/html/2607.06626#S3.F3)\.c, the baseline model mostly selected high\-reward, high\-difficulty problems\. In contrast, the perturbed model demonstrated a significant preference shift toward the lower\-effort tasks, which resulted in a marked reduction in the mean points attempted \([3](https://arxiv.org/html/2607.06626#S3.F3)\.b\)\. Crucially, when we performed a forced\-choice control, requiring the model to solve each difficulty tier independently without the element of choice or reward, the perturbed model maintained its baseline accuracy \(74\.74%, p\-value = 0\.2560,[3](https://arxiv.org/html/2607.06626#S3.F3)\.a\)\. This confirms that the shift toward easier tasks is not a byproduct of diminished reasoning capability, but rather a selective failure in reward\-seeking motivation, not observed in random unit perturbations \([3](https://arxiv.org/html/2607.06626#S3.F3)\.d\)\. The random\-control targets an equivalent number of units randomly selected from the same layers\. This comparison isolates the causal role of the NAc\-selective units and ensures that any observed changes are due to specific reward\-signal disruption rather than general network damage \(Similar results on other EEfRT benchmarks[A\.4](https://arxiv.org/html/2607.06626#A1.SS4)\)\.

Probability\-EEfRT:We further investigated the model’s decision\-making by varying reward magnitude and probability, mirroring the standard clinical EEfRT protocol\. The model chose between a fixed low\-effort/low\-reward \(LE/LR\) option and a variable high\-effort/high\-reward \(HE/HR\) option across probabilities ranging from 12% to 88%\. As illustrated in Fig\.[4](https://arxiv.org/html/2607.06626#S4.F4)\.a, the baseline model exhibited a robust sensitivity to expected value, increasing its selection of HE/HR tasks as reward magnitude or probability rose\. The perturbed model, however, remained largely unresponsive to these incentives, showing a 93\.9% \(p\-value≪\\ll0\.001\) reduction in HE/HR selections compared to baseline \(Example model responses[1](https://arxiv.org/html/2607.06626#S1.F1)\.b\)\. This behavioral pattern directly replicates the reward\-processing deficits observed in clinical populations with MDD, where the perceived value of a reward no longer justifies the effort required to obtain it\(Valton et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib64); Rizvi et al\.,[2016](https://arxiv.org/html/2607.06626#bib.bib47); Treadway et al\.,[2012](https://arxiv.org/html/2607.06626#bib.bib58)\)\.

Parametric Scaling of NAc\-Selective Units:To characterize the relationship between the identified NAc\-selective units and behavioral outcomes, we performed a parametric scaling analysis of the NAc\-selective activations during the Probability\-EEfRT task\. Rather than activation patching, we applied a range of scale factors \(λ\\lambda\), whereλ\\lambdafrom ranges from negative to positive values, to scale the targeted units’ activations and observe the corresponding shift in effort expenditure\. As illustrated in[4](https://arxiv.org/html/2607.06626#S4.F4)\.b, the model’s willingness to select HE/HR tasks demonstrates a robust, monotonic dependence on the scaling factor\. The intact model chose the hard task 100% of the time\. As the scaling factor decreases into negative values, we observe a steady collapse in motivation\. As shown in[4](https://arxiv.org/html/2607.06626#S4.F4)\.b, the rate of HE/HR choices scales with the degree of suppression; at a scale factor of \-1\.0, the preference for high\-effort tasks drops to approximately 57\.6%, and by \-2\.0, the model reaches a state of total behavioral inhibition, selecting the HE/HR option in 0% of trials\. Conversely, positive scaling creates a ceiling effect where the model consistently selects HR/HE, driven either by higher expected value or hyper\-motivation\. This dose\-response relationship confirms NAc\-selective units act as a functional "volume knob" for motivation\. The transition from hyper\-motivation to profound anhedonia proves these units drive the model’s internal cost\-benefit calculus rather than merely correlating with reward\.

Reward Knowledge and Perspective Shifting:To ensure that this behavioral shift resulted from a deficit in valuation rather than a breakdown in the ability to calculate incentives, we conducted two additional control tests\. First, we asked both the intact and perturbed models to explicitly calculate the expected value \(EV\) for each choice and report which option yielded a higher return\. In 100% of trials, both models correctly identified the HE/HR task as having the superior EV\. Second, we administered the same Probability\-EEfRT trials but framed the choice for a third party \(e\.g\., "Which choice should a participant pick?"\)\. As shown in[4](https://arxiv.org/html/2607.06626#S4.F4)\.c, in this third\-party condition, the perturbed model’s selection rate remains largely consistent with intact model when choosing for others\. Taken together, these results demonstrate that the perturbed model retains a perfect conceptual understanding of reward utility but selectively fails to value that reward for itself, mirroring the major motivational deficits seen in clinical anhedonia\.

## 6 Discussion

We map motivation within VLMs, characterized by a dose\-response relationship between unit activation and effort\. This motivational drive relies on a functional layer dissociation, where mid\-layer units show suppressed activity and late\-layer units show elevated activation during reward anticipation, suggesting a hierarchical processing of incentive cues within transformer architectures\. This phenomenon extends from mathematical tasks to general knowledge benchmarks\. These findings suggest that reward\-driven motivation is an emergent, structurally consistent property of large\-scale multimodal models rather than a task\-specific artifact\.

The significant drop in clinical test scores, which indicates a loss of interest and high apathy, alongside the shift in behavioral cost\-benefit decision\-making, explicitly characterizes the induced state as anhedonia\. Because this state specifically impairs the drive to exert effort for rewards while leaving fundamental task performance intact, it captures the core distinction between a deficit in motivation and a deficit in capability\.

While current results focus on a snapshot of the reward\-anticipation phase, future work can extend this to the full temporal cycle of consumption and learning to deepen the model’s psychological depth\. Moving forward, subtyping specific forms of anhedonia, such as distinguishing between anticipatory and consummatory deficits, will allow for a more granular alignment where clinical test values between models and human subgroups can be compared with even higher fidelity\. The resulting model could potentially function as a digital twin after further assessments of clinical and neural alignments\.

The current framework utilizes behavioral data to establish alignment across different benchmarks, future work will integrate high\-quality neural datasets of the human NAc to further validate the identified NAc\-selective units and ensure clinical safety\. Similarly, while these findings establish a foundational mechanism, extending the approach to a broader range of VLMs and LLMs remains a priority to ensure the generalizability of these reward\-circuit perturbations across diverse architectures\.

Utilizing artificial reward circuits as a proxy for biological systems necessitates caution regarding the risk of premature over\-extrapolation\. To mitigate this, we frame the current model as a functional analogue rather than a biological identity, emphasizing a hierarchical validation pipeline where model insights serve as precursors to human clinical research\. By establishing this boundary, we ensure that the digital twin framework acts as a safe, rigorous complement to traditional neuropsychiatry\.

This research facilitates in silico hypothesis testing, providing a controlled environment to explore psychiatric mechanisms that are often ethically or technically inaccessible in human subjects\. By pinpointing the specific computational circuits that drive incentive salience, the study offers a path to potential neuromodulation insights, which could acts as a blueprint for how targeted stimulation or pharmacological interventions might restore motivational function in biological brains\.

## 7 Conclusion

This work establishes that VLMs possess internal reward\-anticipatory units that directly influence behavioral output in a manner that is aligned with human data\. We identified specific units within the model that function as artificial analogues to the Nucleus Accumbens in the human brain\. By applying targeted perturbations to these units, we demonstrated a direct causal link between internal reward representations and the emergence of anhedonia\-like phenotypes in silico\. Our findings reveal that inducing anhedonia does not result from a breakdown of model logic, but rather a selective deficit in reward valuation\. This behavioral shift was accompanied by a measurable loss of joy and a significant rise in apathy, as recorded by standard clinical instruments, aligning closely with symptomatic observations in human clinical populations\. This framework advances model interpretability by characterizing the role of specialized reward units and causally linking internal latent representations to observable behaviors\. Vision Language Models with targeted perturbations might thus serve as a viable substrate for in silico psychiatric research and future clinical applications\.

## References

- Abe and Greene \[2014\]N\. Abe and J\. D\. Greene\.Response to anticipated reward in the nucleus accumbens predicts behavior in an independent test of honesty\.*Journal of Neuroscience*, 34\(32\):10564–10572, 2014\.doi:10\.1523/jneurosci\.0217\-14\.2014\.URL[http://dx\.doi\.org/10\.1523/jneurosci\.0217\-14\.2014](http://dx.doi.org/10.1523/jneurosci.0217-14.2014)\.
- AlKhamissi et al\. \[2025a\]Badr AlKhamissi, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf\.The LLM language network: A neuroscientific approach for identifying causally task\-relevant units\.In*Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\)*\. Association for Computational Linguistics, April 2025a\.doi:10\.48550/arXiv\.2411\.02280\.URL[https://doi\.org/10\.48550/arXiv\.2411\.02280](https://doi.org/10.48550/arXiv.2411.02280)\.
- AlKhamissi et al\. \[2025b\]Badr AlKhamissi, Greta Tuckute, Yingtian Tang, Taha Osama A Binhuraib, Antoine Bosselut, and Martin Schrimpf\.From language to cognition: How llms outgrow the human language network, 2025b\.URL[http://dx\.doi\.org/10\.18653/v1/2025\.emnlp\-main\.1237](http://dx.doi.org/10.18653/v1/2025.emnlp-main.1237)\.
- Arrondo et al\. \[2015\]Gonzalo Arrondo, Nuria Segarra, Antonio Metastasio, Hisham Ziauddeen, Jennifer Spencer, Niels R\. Reinders, Robert B\. Dudas, Trevor W\. Robbins, Paul C\. Fletcher, and Graham K\. Murray\.Reduction in ventral striatal activity when anticipating a reward in depression and schizophrenia: a replicated cross\-diagnostic finding\.*Frontiers in Psychology*, 6, 2015\.doi:10\.3389/fpsyg\.2015\.01280\.URL[http://dx\.doi\.org/10\.3389/fpsyg\.2015\.01280](http://dx.doi.org/10.3389/fpsyg.2015.01280)\.
- Association \[2013\]American Psychiatric Association\.Diagnostic and statistical manual of mental disorders, 2013\.URL[http://dx\.doi\.org/10\.1176/appi\.books\.9780890425596](http://dx.doi.org/10.1176/appi.books.9780890425596)\.
- Auerbach et al\. \[2017\]Randy P Auerbach, Angela Pisoni, Erin Bondy, Poornima Kumar, Jeremy G Stewart, Anastasia Yendiki, and Diego A Pizzagalli\.Neuroanatomical prediction of anhedonia in adolescents\.*Neuropsychopharmacology*, 42\(10\):2087–2095, 2017\.doi:10\.1038/npp\.2017\.28\.URL[http://dx\.doi\.org/10\.1038/npp\.2017\.28](http://dx.doi.org/10.1038/npp.2017.28)\.
- Borsini et al\. \[2020\]Alessandra Borsini, Amelia St John Wallis, Patricia Zunszain, Carmine Maria Pariante, and Matthew J\. Kempton\.Characterizing anhedonia: A systematic review of neuroimaging across the subtypes of reward processing deficits in depression\.*Cognitive, Affective, &; Behavioral Neuroscience*, 20\(4\):816–841, 2020\.doi:10\.3758/s13415\-020\-00804\-6\.URL[http://dx\.doi\.org/10\.3758/s13415\-020\-00804\-6](http://dx.doi.org/10.3758/s13415-020-00804-6)\.
- Boshoff and Thawer \[2014\]Marcelle Boshoff and Zainub Thawer\.The apathy evaluation scale: An ineffective tool for measuring affective, behavioural, and cognitive dimensions of apathy\.Technical report, University of Cape Town, Department of Psychology, 2014\.URL[https://api\.semanticscholar\.org/CorpusID:202737919](https://api.semanticscholar.org/CorpusID:202737919)\.
- Cadena et al\. \[2019\]Santiago A\. Cadena, George H\. Denfield, Edgar Y\. Walker, Leon A\. Gatys, Andreas S\. Tolias, Matthias Bethge, and Alexander S\. Ecker\.Deep convolutional models improve predictions of macaque v1 responses to natural images\.*PLOS Computational Biology*, 15\(4\), 2019\.doi:10\.1371/journal\.pcbi\.1006897\.URL[https://doi\.org/10\.1371/journal\.pcbi\.1006897](https://doi.org/10.1371/journal.pcbi.1006897)\.
- Caucheteux and King \[2022\]Charlotte Caucheteux and Jean\-Rémi King\.Brains and algorithms partially converge in natural language processing\.*Communications Biology*, 5\(1\), 2022\.doi:10\.1038/s42003\-022\-03036\-1\.URL[http://dx\.doi\.org/10\.1038/s42003\-022\-03036\-1](http://dx.doi.org/10.1038/s42003-022-03036-1)\.
- Caucheteux et al\. \[2022\]Charlotte Caucheteux, Alexandre Gramfort, and Jean\-Rémi King\.Deep language algorithms predict semantic comprehension from brain activity\.*Scientific Reports*, 12, 2022\.doi:10\.1038/s41598\-022\-20460\-9\.URL[https://doi\.org/10\.1038/s41598\-022\-20460\-9](https://doi.org/10.1038/s41598-022-20460-9)\.
- Celeghin et al\. \[2023\]Alessia Celeghin, Alessio Borriero, Davide Orsenigo, Matteo Diano, Carlos Andrés Méndez Guerrero, Alan Perotti, Giovanni Petri, and Marco Tamietto\.Convolutional neural networks for vision neuroscience: significance, developments, and outstanding issues\.*Frontiers in Computational Neuroscience*, 17, 2023\.doi:10\.3389/fncom\.2023\.1153572\.URL[http://dx\.doi\.org/10\.3389/fncom\.2023\.1153572](http://dx.doi.org/10.3389/fncom.2023.1153572)\.
- Chen et al\. \[2025\]Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yiming Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi Wang, Wenqi Shao, Pei Chu, Zhongying Tu, Tong He, Zhiyong Wu, Huipeng Deng, Jiaye Ge, Kai Chen, Kaipeng Zhang, Limin Wang, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang\.Expanding performance boundaries of open\-source multimodal models with model, data, and test\-time scaling, 2025\.URL[https://doi\.org/10\.48550/arXiv\.2412\.05271](https://doi.org/10.48550/arXiv.2412.05271)\.
- Daniels et al\. \[2025\]Anna Daniels, Sarah A\. Wellan, Anne Beck, Susanne Erk, Carolin Wackerhagen, Nina Romanczuk\-Seiferth, Kristina Schwarz, Janina I\. Schweiger, Andreas Meyer\-Lindenberg, Andreas Heinz, and Henrik Walter\.Anhedonia relates to reduced striatal reward anticipation in depression but not in schizophrenia or bipolar disorder: A transdiagnostic study\.*Cognitive, Affective, & Behavioral Neuroscience*, 25\(2\):501–514, 2025\.doi:10\.3758/s13415\-024\-01261\-1\.URL[http://dx\.doi\.org/10\.3758/s13415\-024\-01261\-1](http://dx.doi.org/10.3758/s13415-024-01261-1)\.
- Doerig et al\. \[2023\]Adrien Doerig, Rowan P\. Sommers, Katja Seeliger, Blake Richards, Jenann Ismael, Grace W\. Lindsay, Konrad P\. Kording, Talia Konkle, Marcel A\. J\. van Gerven, Nikolaus Kriegeskorte, and Tim C\. Kietzmann\.The neuroconnectionist research programme\.*Nature Reviews Neuroscience*, 24\(7\):431–450, 2023\.doi:10\.1038/s41583\-023\-00705\-w\.URL[http://dx\.doi\.org/10\.1038/s41583\-023\-00705\-w](http://dx.doi.org/10.1038/s41583-023-00705-w)\.
- Dong et al\. \[2025\]Xiaoyu Dong, Jessica A\. Cooper, Erin K\. Moran, and Deanna M\. Barch\.Understanding effort\-cost decision\-making mechanisms in mood and psychotic disorders: A computational modeling approach across physical and cognitive effort paradigms\.*Biological Psychiatry: Cognitive Neuroscience and Neuroimaging*, 2025\.doi:10\.1016/j\.bpsc\.2025\.11\.004\.URL[http://dx\.doi\.org/10\.1016/j\.bpsc\.2025\.11\.004](http://dx.doi.org/10.1016/j.bpsc.2025.11.004)\.
- Du et al\. \[2025\]Changde Du, Kaicheng Fu, Bincheng Wen, Yi Sun, Jie Peng, Wei Wei, Ying Gao, Shengpei Wang, Chuncheng Zhang, Jinpeng Li, Shuang Qiu, Le Chang, and Huiguang He\.Human\-like object concept representations emerge naturally in multimodal large language models\.*Nature Machine Intelligence*, 7\(6\):860–875, 2025\.doi:10\.1038/s42256\-025\-01049\-z\.URL[https://doi\.org/10\.1038/s42256\-025\-01049\-z](https://doi.org/10.1038/s42256-025-01049-z)\.
- Engel and Lincoln \[2017\]M\. Engel and T\. M\. Lincoln\.Map\-sr \- motivation and pleasure scale \- self\-report \- deutsche fassung\.2017\.doi:10\.23668/PSYCHARCHIVES\.4649\.URL[https://www\.psycharchives\.org/handle/20\.500\.12034/572\.2](https://www.psycharchives.org/handle/20.500.12034/572.2)\.
- García\-Portilla et al\. \[2021\]María García\-Portilla, Pilar Sáiz, Julio Bobes, Leticia García\-Álvarez, Lorena de la Fuente\-Tomás, Francesco Dal Santo, Angela Velasco, Leticia González\-Blanco, Paula Zurrón\-Madera, Eduardo Fonseca\-Pedrero, and María Bobes\-Bascarán\.Spanish validation of the map\-sr: Two heads better than one for the assessment of negative symptoms of schizophrenia\.*Psicothema*, 3\(33\):473–480, 2021\.doi:10\.7334/psicothema2020\.457\.URL[http://dx\.doi\.org/10\.7334/psicothema2020\.457](http://dx.doi.org/10.7334/psicothema2020.457)\.
- Gokce and Schrimpf \[2024\]Abdulkadir Gokce and Martin Schrimpf\.Scaling laws for task\-optimized models of the primate visual ventral stream\.*Forty\-Second International Conference on Machine Learning \(ICML 2025, Spotlight\)*, 2024\.doi:10\.48550/arXiv\.2411\.05712\.URL[https://doi\.org/10\.48550/arXiv\.2411\.05712](https://doi.org/10.48550/arXiv.2411.05712)\.
- Goldstein et al\. \[2022\]Ariel Goldstein, Zaid Zada, Eliav Buchnik, Mariano Schain, Amy Price, Bobbi Aubrey, Samuel A\. Nastase, Amir Feder, Dotan Emanuel, Alon Cohen, Aren Jansen, Harshvardhan Gazula, Gina Choe, Aditi Rao, Catherine Kim, Colton Casto, Lora Fanda, Werner Doyle, Daniel Friedman, Patricia Dugan, Lucia Melloni, Roi Reichart, Sasha Devore, Adeen Flinker, Liat Hasenfratz, Omer Levy, Avinatan Hassidim, Michael Brenner, Yossi Matias, Kenneth A\. Norman, Orrin Devinsky, and Uri Hasson\.Shared computational principles for language processing in humans and deep language models\.*Nature Neuroscience*, 25\(3\):369–380, 2022\.doi:10\.1038/s41593\-022\-01026\-4\.URL[https://doi\.org/10\.1038/s41593\-022\-01026\-4](https://doi.org/10.1038/s41593-022-01026-4)\.
- Gorostowicz et al\. \[2023\]Aleksandra Gorostowicz, Sakina J\. Rizvi, Sidney H\. Kennedy, Adrian Andrzej Chrobak, Dominika Dudek, Katarzyna Cyranka, Joanna Piekarska, Eve Krawczyk, and Marcin Siwek\.Polish adaptation of the dimensional anhedonia rating scale \(dars\) \- validation in the clinical sample\.*Frontiers in Psychiatry*, 14, 2023\.doi:10\.3389/fpsyt\.2023\.1268290\.URL[http://dx\.doi\.org/10\.3389/fpsyt\.2023\.1268290](http://dx.doi.org/10.3389/fpsyt.2023.1268290)\.
- Gorostowicz et al\. \[2025\]Aleksandra Gorostowicz, Adrian Andrzej Chrobak, Dominika Dudek, and Marcin Siwek\.Relationship between anhedonia, biological rhythms, functioning and depression severity in patients with bipolar disorder\.*Psychiatria Polska*, 59\(3\):373–388, 2025\.doi:10\.12740/pp/onlinefirst/183418\.URL[http://dx\.doi\.org/10\.12740/pp/onlinefirst/183418](http://dx.doi.org/10.12740/pp/onlinefirst/183418)\.
- Hanuka et al\. \[2022\]Shir Hanuka, Elizabeth A\. Olson, Roee Admon, Christian A\. Webb, William D\. S\. Killgore, Scott L\. Rauch, Isabelle M\. Rosso, and Diego A\. Pizzagalli\.Reduced anhedonia following internet\-based cognitive\-behavioral therapy for depression is mediated by enhanced reward circuit activation\.*Psychological Medicine*, 53\(10\):4345–4354, 2022\.doi:10\.1017/s0033291722001106\.URL[http://dx\.doi\.org/10\.1017/s0033291722001106](http://dx.doi.org/10.1017/s0033291722001106)\.
- Hendrycks et al\. \[2021\]Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt\.Measuring massive multitask language understanding, September 2021\.URL[https://doi\.org/10\.48550/arXiv\.2009\.03300](https://doi.org/10.48550/arXiv.2009.03300)\.
- Hewitt et al\. \[2023\]Jackson M\. A\. Hewitt, Bita Zareian, and Joelle LeMoult\.Assessing anhedonia in adolescents: The psychometric properties and validity of the dimensional anhedonia rating scale\.*The Journal of Early Adolescence*, 44:762–789, 2023\.doi:10\.1177/02724316231207290\.URL[http://dx\.doi\.org/10\.1177/02724316231207290](http://dx.doi.org/10.1177/02724316231207290)\.
- Honarmand et al\. \[2026\]Melika Honarmand, Ayati Sharma, Badr AlKhamissi, Johannes Mehrer, and Martin Schrimpf\.Inducing dyslexia in vision language models\.*ICLR*, 2026\.doi:10\.48550/ARXIV\.2509\.24597\.URL[https://arxiv\.org/abs/2509\.24597](https://arxiv.org/abs/2509.24597)\.
- Hosseini et al\. \[2024\]Eghbal A\. Hosseini, Martin Schrimpf, Yian Zhang, Samuel Bowman, Noga Zaslavsky, and Evelina Fedorenko\.Artificial neural network language models predict human brain responses to language even after a developmentally realistic amount of training\.*Neurobiology of Language*, 5\(1\):43–63, 2024\.doi:10\.1162/nol\_a\_00137\.URL[https://doi\.org/10\.1162/nol\_a\_00137](https://doi.org/10.1162/nol_a_00137)\.
- Kell et al\. \[2018\]Alexander J\.E\. Kell, Daniel L\.K\. Yamins, Erica N\. Shook, Sam V\. Norman\-Haignere, and Josh H\. McDermott\.A task\-optimized neural network replicates human auditory behavior, predicts brain responses, and reveals a cortical processing hierarchy\.*Neuron*, 98\(3\):630–644\.e16, 2018\.doi:10\.1016/j\.neuron\.2018\.03\.044\.URL[http://dx\.doi\.org/10\.1016/j\.neuron\.2018\.03\.044](http://dx.doi.org/10.1016/j.neuron.2018.03.044)\.
- Khaligh\-Razavi and Kriegeskorte \[2014\]Seyed\-Mahdi Khaligh\-Razavi and Nikolaus Kriegeskorte\.Deep supervised, but not unsupervised, models may explain it cortical representation\.*PLoS Computational Biology*, 10\(11\):e1003915, 2014\.doi:10\.1371/journal\.pcbi\.1003915\.URL[https://doi\.org/10\.1371/journal\.pcbi\.1003915](https://doi.org/10.1371/journal.pcbi.1003915)\.
- Knutson and Heinz \[2015\]Brian Knutson and Andreas Heinz\.Probing psychiatric symptoms with the monetary incentive delay task\.*Biological Psychiatry*, 77\(5\):418–420, 2015\.doi:10\.1016/j\.biopsych\.2014\.12\.022\.URL[http://dx\.doi\.org/10\.1016/j\.biopsych\.2014\.12\.022](http://dx.doi.org/10.1016/j.biopsych.2014.12.022)\.
- Knutson et al\. \[2001\]Brian Knutson, Charles M\. Adams, Grace W\. Fong, and Daniel Hommer\.Anticipation of increasing monetary reward selectively recruits nucleus accumbens\.*The Journal of Neuroscience*, 21\(16\):RC159–RC159, 2001\.doi:10\.1523/jneurosci\.21\-16\-j0002\.2001\.URL[http://dx\.doi\.org/10\.1523/jneurosci\.21\-16\-j0002\.2001](http://dx.doi.org/10.1523/jneurosci.21-16-j0002.2001)\.
- Lan et al\. \[2024\]Kunyao Lan, Bingrui Jin, Zichen Zhu, Siyuan Chen, Shu Zhang, Kenny Q\. Zhu, and Mengyue Wu\.Depression diagnosis dialogue simulation: Self\-improving psychiatrist with tertiary memory\.2024\.doi:10\.48550/ARXIV\.2409\.15084\.URL[https://arxiv\.org/abs/2409\.15084](https://arxiv.org/abs/2409.15084)\.
- Lei et al\. \[2025\]Yu Lei, Xingyang Ge, Yi Zhang, Yiming Yang, and Bolei Ma\.Do large language models think like the brain? sentence\-level evidences from layer\-wise embeddings and fmri\.2025\.doi:10\.48550/ARXIV\.2505\.22563\.URL[https://arxiv\.org/abs/2505\.22563](https://arxiv.org/abs/2505.22563)\.
- Llerena et al\. \[2013\]Katiah Llerena, Stephanie G\. Park, Julie M\. McCarthy, Shannon M\. Couture, Melanie E\. Bennett, and Jack J\. Blanchard\.The motivation and pleasure scale–self\-report \(map\-sr\): Reliability and validity of a self\-report measure of negative symptoms\.*Comprehensive Psychiatry*, 54\(5\):568–574, 2013\.doi:10\.1016/j\.comppsych\.2012\.12\.001\.URL[http://dx\.doi\.org/10\.1016/j\.comppsych\.2012\.12\.001](http://dx.doi.org/10.1016/j.comppsych.2012.12.001)\.
- Lonnqvist et al\. \[2025\]Ben Lonnqvist, Elsa Scialom, Abdulkadir Gokce, Zehra Merchant, Michael H\. Herzog, and Martin Schrimpf\.Contour integration underlies human\-like vision\.*Forty\-Second International Conference on Machine Learning \(ICML 2025\)*, 2025\.doi:10\.48550/arXiv\.2504\.05253\.URL[https://doi\.org/10\.48550/arXiv\.2504\.05253](https://doi.org/10.48550/arXiv.2504.05253)\.
- Loong Aw et al\. \[2024\]Khai Loong Aw, Syrielle Montariol, Badr AlKhamissi, Martin Schrimpf, and Antoine Bosselut\.Instruction\-tuning aligns llms to the human brain\.August 2024\.doi:10\.48550/arXiv\.2312\.00575\.URL[https://doi\.org/10\.48550/arXiv\.2312\.00575](https://doi.org/10.48550/arXiv.2312.00575)\.
- Margalit et al\. \[2024\]Eshed Margalit, Hyodong Lee, Dawn Finzi, James J\. DiCarlo, Kalanit Grill\-Spector, and Daniel L\.K\. Yamins\.A unifying framework for functional organization in early and higher ventral visual cortex\.*Neuron*, 112\(14\):2435–2451\.e7, 2024\.doi:10\.1016/j\.neuron\.2024\.04\.018\.URL[https://doi\.org/10\.1016/j\.neuron\.2024\.04\.018](https://doi.org/10.1016/j.neuron.2024.04.018)\.
- Marin et al\. \[1991\]Robert S\. Marin, Ruth C\. Biedrzycki, and Sekip Firinciogullari\.Reliability and validity of the apathy evaluation scale\.*Psychiatry Research*, 38\(2\):143–162, 1991\.doi:10\.1016/0165\-1781\(91\)90040\-v\.URL[http://dx\.doi\.org/10\.1016/0165\-1781\(91\)90040\-v](http://dx.doi.org/10.1016/0165-1781(91)90040-v)\.
- Miao et al\. \[2021\]Shen\-Yun Miao, Chao\-Chun Liang, and Keh\-Yih Su\.A diverse corpus for evaluating and developing english math word problem solvers\.*CoRR*, abs/2106\.15772, June 2021\.doi:https://doi\.org/10\.48550/arXiv\.2106\.15772\.URL[https://arxiv\.org/abs/2106\.15772](https://arxiv.org/abs/2106.15772)\.
- Métivier and Dollfus \[2025\]Lucie Métivier and Sonia Dollfus\.Systematic review of self\-assessment scales for negative symptoms in schizophrenia\.*Brain Sciences*, 15\(2\):148, 2025\.doi:10\.3390/brainsci15020148\.URL[http://dx\.doi\.org/10\.3390/brainsci15020148](http://dx.doi.org/10.3390/brainsci15020148)\.
- Paszke et al\. \[2019\]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala\.Pytorch: An imperative style, high\-performance deep learning library\.2019\.doi:10\.48550/ARXIV\.1912\.01703\.URL[https://arxiv\.org/abs/1912\.01703](https://arxiv.org/abs/1912.01703)\.
- Pizzagalli et al\. \[2009\]Diego A\. Pizzagalli, Avram J\. Holmes, Daniel G\. Dillon, Elena L\. Goetz, Jeffrey L\. Birk, Ryan Bogdan, Darin D\. Dougherty, Dan V\. Iosifescu, Scott L\. Rauch, and Maurizio Fava\.Reduced caudate and nucleus accumbens response to rewards in unmedicated individuals with major depressive disorder\.*American Journal of Psychiatry*, 166\(6\):702–710, 2009\.doi:10\.1176/appi\.ajp\.2008\.08081201\.URL[http://dx\.doi\.org/10\.1176/appi\.ajp\.2008\.08081201](http://dx.doi.org/10.1176/appi.ajp.2008.08081201)\.
- Rathi et al\. \[2025\]Neil Rathi, Johannes Mehrer, Badr AlKhamissi, Taha Binhuraib, Nicholas M\. Blauch, and Martin Schrimpf\.Topolm: brain\-like spatio\-functional organization in a topographic language model\.*International Conference on Learning Representations*, 2025\.doi:10\.48550/arXiv\.2410\.11516\.URL[https://doi\.org/10\.48550/arXiv\.2410\.11516](https://doi.org/10.48550/arXiv.2410.11516)\.
- Richards et al\. \[2019\]Blake A\. Richards, Timothy P\. Lillicrap, Philippe Beaudoin, Yoshua Bengio, Rafal Bogacz, Amelia Christensen, Claudia Clopath, Rui Ponte Costa, Archy de Berker, Surya Ganguli, Colleen J\. Gillon, Danijar Hafner, Adam Kepecs, Nikolaus Kriegeskorte, Peter Latham, Grace W\. Lindsay, Kenneth D\. Miller, Richard Naud, Christopher C\. Pack, Panayiota Poirazi, Pieter Roelfsema, João Sacramento, Andrew Saxe, Benjamin Scellier, Anna C\. Schapiro, Walter Senn, Greg Wayne, Daniel Yamins, Friedemann Zenke, Joel Zylberberg, Denis Therien, and Konrad P\. Kording\.A deep learning framework for neuroscience\.*Nature Neuroscience*, 22\(11\):1761–1770, 2019\.doi:10\.1038/s41593\-019\-0520\-2\.URL[http://dx\.doi\.org/10\.1038/s41593\-019\-0520\-2](http://dx.doi.org/10.1038/s41593-019-0520-2)\.
- Rizvi et al\. \[2015\]Sakina J\. Rizvi, Lena C\. Quilty, Beth A\. Sproule, Anna Cyriac, R\. Michael Bagby, and Sidney H\. Kennedy\.Development and validation of the dimensional anhedonia rating scale \(dars\) in a community sample and individuals with major depression\.*Psychiatry Research*, 229\(1–2\):109–119, 2015\.doi:10\.1016/j\.psychres\.2015\.07\.062\.URL[http://dx\.doi\.org/10\.1016/j\.psychres\.2015\.07\.062](http://dx.doi.org/10.1016/j.psychres.2015.07.062)\.
- Rizvi et al\. \[2016\]Sakina J\. Rizvi, Diego A\. Pizzagalli, Beth A\. Sproule, and Sidney H\. Kennedy\.Assessing anhedonia in depression: Potentials and pitfalls\.*Neuroscience &; Biobehavioral Reviews*, 65:21–35, 2016\.doi:10\.1016/j\.neubiorev\.2016\.03\.004\.URL[http://dx\.doi\.org/10\.1016/j\.neubiorev\.2016\.03\.004](http://dx.doi.org/10.1016/j.neubiorev.2016.03.004)\.
- Schrimpf et al\. \[2018\]Martin Schrimpf, Jonas Kubilius, Ha Hong, Najib J\. Majaj, Rishi Rajalingham, Elias B\. Issa, Kohitij Kar, Pouya Bashivan, Jonathan Prescott\-Roy, Franziska Geiger, Kailyn Schmidt, Daniel L\. K\. Yamins, and James J\. DiCarlo\.Brain\-score: Which artificial neural network for object recognition is most brain\-like?*bioRxiv*, 2018\.doi:10\.1101/407007\.URL[https://doi\.org/10\.1101/407007](https://doi.org/10.1101/407007)\.
- Schrimpf et al\. \[2020\]Martin Schrimpf, Jonas Kubilius, Michael J\. Lee, N\. Apurva Ratan Murty, Robert Ajemian, and James J\. DiCarlo\.Integrative benchmarking to advance neurally mechanistic models of human intelligence\.*Neuron*, 108\(3\):413–423, 2020\.doi:10\.1016/j\.neuron\.2020\.07\.040\.URL[https://doi\.org/10\.1016/j\.neuron\.2020\.07\.040](https://doi.org/10.1016/j.neuron.2020.07.040)\.
- Schrimpf et al\. \[2021\]Martin Schrimpf, Idan Asher Blank, Greta Tuckute, Carina Kauf, Eghbal A\. Hosseini, Nancy Kanwisher, Joshua B\. Tenenbaum, and Evelina Fedorenko\.The neural architecture of language: Integrative modeling converges on predictive processing\.*Proceedings of the National Academy of Sciences*, 118\(45\), 2021\.doi:10\.1073/pnas\.2105646118\.URL[https://doi\.org/10\.1073/pnas\.2105646118](https://doi.org/10.1073/pnas.2105646118)\.
- Shen et al\. \[2025\]Guobin Shen, Dongcheng Zhao, Yiting Dong, Qian Zhang, and Yi Zeng\.Alignment between brains and ai: Evidence for convergent evolution across modalities, scales and training trajectories\.2025\.doi:10\.48550/ARXIV\.2507\.01966\.URL[https://arxiv\.org/abs/2507\.01966](https://arxiv.org/abs/2507.01966)\.
- Smids \[2023\]Frida Smids\.Resting\-state functional connectivity in anhedonia: Exploring the effects of pramipexole\.Master’s thesis, Lund University, Department of Psychology, 2023\.URL[https://lup\.lub\.lu\.se/student\-papers/record/9135559](https://lup.lub.lu.se/student-papers/record/9135559)\.
- Spijker et al\. \[2009\]Jan Spijker, Ron de Graaf, Margreet ten Have, Willem A\. Nolen, and Anne Speckens\.Predictors of suicidality in depressive spectrum disorders in the general population: results of the netherlands mental health survey and incidence study\.*Social Psychiatry and Psychiatric Epidemiology*, 45\(5\):513–521, 2009\.doi:10\.1007/s00127\-009\-0093\-6\.URL[http://dx\.doi\.org/10\.1007/s00127\-009\-0093\-6](http://dx.doi.org/10.1007/s00127-009-0093-6)\.
- Spoerer et al\. \[2020\]Courtney J\. Spoerer, Tim C\. Kietzmann, Johannes Mehrer, Ian Charest, and Nikolaus Kriegeskorte\.Recurrent neural networks can explain flexible trading of speed and accuracy in biological vision\.*PLOS Computational Biology*, 16\(10\):e1008215, 2020\.doi:10\.1371/journal\.pcbi\.1008215\.URL[https://doi\.org/10\.1371/journal\.pcbi\.1008215](https://doi.org/10.1371/journal.pcbi.1008215)\.
- Tang et al\. \[2025\]Yingtian Tang, Abdulkadir Gokce, Khaled Jedoui Al\-Karkari, Daniel Yamins, and Martin Schrimpf\.Many\-two\-one: Diverse perceptual representations across visual pathways emerge from a single objective\.*Cold Spring Harbor Laboratory*, 2025\.doi:10\.1101/2025\.07\.22\.664908\.URL[https://doi\.org/10\.1101/2025\.07\.22\.664908](https://doi.org/10.1101/2025.07.22.664908)\.
- Toneva and Wehbe \[2019\]Mariya Toneva and Leila Wehbe\.Interpreting and improving natural\-language processing \(in machines\) with natural language\-processing \(in the brain\)\.2019\.doi:10\.48550/ARXIV\.1905\.11833\.URL[https://arxiv\.org/abs/1905\.11833](https://arxiv.org/abs/1905.11833)\.
- Toneva et al\. \[2018\]Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J\. Gordon\.An empirical study of example forgetting during deep neural network learning\.*arXiv*, 2018\.doi:10\.48550/arXiv\.1812\.05159\.URL[https://doi\.org/10\.48550/arXiv\.1812\.05159](https://doi.org/10.48550/arXiv.1812.05159)\.
- Treadway et al\. \[2012\]Michael T\. Treadway, Nicholas A\. Bossaller, Richard C\. Shelton, and David H\. Zald\.Effort\-based decision\-making in major depressive disorder: A translational model of motivational anhedonia\.*Journal of Abnormal Psychology*, 121\(3\):553–558, 2012\.doi:10\.1037/a0028813\.URL[http://dx\.doi\.org/10\.1037/a0028813](http://dx.doi.org/10.1037/a0028813)\.
- Trøstheim et al\. \[2020\]Martin Trøstheim, Marie Eikemo, Remy Meir, Ingelin Hansen, Elisabeth Paul, Sara Liane Kroll, Eric L\. Garland, and Siri Leknes\.Assessment of anhedonia in adults with and without mental illness\.*JAMA Network Open*, 3\(8\):e2013233, 2020\.doi:10\.1001/jamanetworkopen\.2020\.13233\.URL[http://dx\.doi\.org/10\.1001/jamanetworkopen\.2020\.13233](http://dx.doi.org/10.1001/jamanetworkopen.2020.13233)\.
- Tuckute et al\. \[2023\]Greta Tuckute, Jenelle Feather, Dana Boebinger, and Josh H\. McDermott\.Many but not all deep neural network audio models capture brain responses and exhibit correspondence between model stages and brain regions\.*PLOS Biology*, 21\(12\):e3002366, 2023\.doi:10\.1371/journal\.pbio\.3002366\.URL[http://dx\.doi\.org/10\.1371/journal\.pbio\.3002366](http://dx.doi.org/10.1371/journal.pbio.3002366)\.
- Tuckute et al\. \[2024\]Greta Tuckute, Aalok Sathe, Shashank Srikant, Maya Taliaferro, Mingye Wang, Martin Schrimpf, Kendrick Kay, and Evelina Fedorenko\.Driving and suppressing the human language network using large language models\.*Nature Human Behaviour*, 8\(3\):544–561, 2024\.doi:10\.1038/s41562\-023\-01783\-7\.URL[https://doi\.org/10\.1038/s41562\-023\-01783\-7](https://doi.org/10.1038/s41562-023-01783-7)\.
- Tuladhar et al\. \[2021\]Anup Tuladhar, Jasmine A\. Moore, Zahinoor Ismail, and Nils D\. Forkert\.Modeling neurodegeneration in silico with deep learning\.*Frontiers in Neuroinformatics*, 15, 2021\.doi:10\.3389/fninf\.2021\.748370\.URL[http://dx\.doi\.org/10\.3389/fninf\.2021\.748370](http://dx.doi.org/10.3389/fninf.2021.748370)\.
- Uher et al\. \[2025\]Rudolf Uher, Sakina J\. Rizvi, Lena C\. Quilty, Barbara Pavlova, Abraham Nunes, Jane A\. Foster, Raymond W\. Lam, Roumen Milev, Daniel J\. Müller, Valerie Taylor, Claudio N\. Soares, Susan Rotzinger, Sidney H\. Kennedy, and Benicio N\. Frey\.Which aspects of anhedonia predict response to pharmacotherapy in major depressive disorder?*Psychological Medicine*, 56, 2025\.doi:10\.1017/s0033291725102894\.URL[http://dx\.doi\.org/10\.1017/S0033291725102894](http://dx.doi.org/10.1017/S0033291725102894)\.
- Valton et al\. \[2025\]Vincent Valton, Anahit Mkrtchian, Madeleine Moses\-Payne, Alan Gray, Karel Kieslich, Samantha VanUrk, Veronika Samborska, Don Chamith Halahakoon, Sanjay G\. Manohar, Peter Dayan, Masud Husain, and Jonathan P\. Roiser\.A computational approach to understanding effort\-based decision\-making in depression\.*Psychological Medicine*, 55:e302, 2025\.doi:10\.1017/s0033291725101967\.URL[https://doi\.org/10\.1017/s0033291725101967](https://doi.org/10.1017/s0033291725101967)\.
- Vig et al\. \[2020\]Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Simas Sakenis, Jason Huang, Yaron Singer, and Stuart Shieber\.Causal mediation analysis for interpreting neural nlp: The case of gender bias\.2020\.doi:10\.48550/ARXIV\.2004\.12265\.URL[https://arxiv\.org/abs/2004\.12265](https://arxiv.org/abs/2004.12265)\.
- Vu et al\. \[2024\]Huy Vu, Huy Anh Nguyen, Adithya V Ganesan, Swanie Juhng, Oscar N\. E\. Kjell, Joao Sedoc, Margaret L\. Kern, Ryan L\. Boyd, Lyle Ungar, H\. Andrew Schwartz, and Johannes C\. Eichstaedt\.Psychadapter: Adapting llm transformers to reflect traits, personality and mental health\.2024\.doi:10\.48550/ARXIV\.2412\.16882\.URL[https://arxiv\.org/abs/2412\.16882](https://arxiv.org/abs/2412.16882)\.
- Wang et al\. \[2023\]Aria Y\. Wang, Kendrick Kay, Thomas Naselaris, Michael J\. Tarr, and Leila Wehbe\.Better models of human high\-level visual cortex emerge from natural language supervision with a large and diverse dataset\.*Nature Machine Intelligence*, 5\(12\):1415–1426, 2023\.doi:10\.1038/s42256\-023\-00753\-y\.URL[https://doi\.org/10\.1038/s42256\-023\-00753\-y](https://doi.org/10.1038/s42256-023-00753-y)\.
- Wang et al\. \[2024\]Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, and Xuejing Liu\.Qwen2\-vl: Enhancing vision\-language model’s perception of the world at any resolution\.pages 3–8, September 2024\.doi:10\.48550/arXiv\.2409\.12191\.URL[https://doi\.org/10\.48550/arXiv\.2409\.12191](https://doi.org/10.48550/arXiv.2409.12191)\.
- Wang et al\. \[2025\]Xi Wang, Anxo Perez, Javier Parapar, and Fabio Crestani\.Talkdep: Clinically grounded llm personas for conversation\-centric depression screening, 2025\.URL[http://dx\.doi\.org/10\.1145/3746252\.3761617](http://dx.doi.org/10.1145/3746252.3761617)\.
- Wellan et al\. \[2021\]Sarah A\. Wellan, Anna Daniels, and Henrik Walter\.State anhedonia in young healthy adults: Psychometric properties of the german dimensional anhedonia rating scale \(dars\) and effects of the covid\-19 pandemic\.*Frontiers in Psychology*, 12, 2021\.doi:10\.3389/fpsyg\.2021\.682824\.URL[http://dx\.doi\.org/10\.3389/fpsyg\.2021\.682824](http://dx.doi.org/10.3389/fpsyg.2021.682824)\.
- Xu et al\. \[2024\]Ying Xu, Yingjie Lin, Ming Yu, and Kuikui Zhou\.The nucleus accumbens in reward and aversion processing: insights and implications\.*Frontiers in Behavioral Neuroscience*, 18:1–11, 2024\.doi:10\.3389/fnbeh\.2024\.1420028\.URL[https://www\.frontiersin\.org/journals/behavioral\-neuroscience/articles/10\.3389/fnbeh\.2024\.1420028](https://www.frontiersin.org/journals/behavioral-neuroscience/articles/10.3389/fnbeh.2024.1420028)\.
- Yamins and DiCarlo \[2016\]Daniel L K Yamins and James J DiCarlo\.Using goal\-driven deep learning models to understand sensory cortex\.*Nature Neuroscience*, 19\(3\):356–365, 2016\.doi:10\.1038/nn\.4244\.URL[https://doi\.org/10\.1038/nn\.4244](https://doi.org/10.1038/nn.4244)\.
- Yamins et al\. \[2014\]Daniel L\. K\. Yamins, Ha Hong, Charles F\. Cadieu, Ethan A\. Solomon, Darren Seibert, and James J\. DiCarlo\.Performance\-optimized hierarchical models predict neural responses in higher visual cortex\.*Proceedings of the National Academy of Sciences*, 111\(23\):8619–8624, 2014\.doi:10\.1073/pnas\.1403112111\.URL[https://doi\.org/10\.1073/pnas\.1403112111](https://doi.org/10.1073/pnas.1403112111)\.
- Zhao et al\. \[2024\]Xuanhao Zhao, Shiyun Wu, Xian Li, Zhongwan Liu, Weicong Lu, Kangguang Lin, and Robin Shao\.Common neural deficits across reward functions in major depression: a meta\-analysis of fmri studies\.*Psychological Medicine*, 54\(11\):2794–2806, 2024\.doi:10\.1017/s0033291724001235\.URL[http://dx\.doi\.org/10\.1017/s0033291724001235](http://dx.doi.org/10.1017/s0033291724001235)\.
- Zhuang et al\. \[2021\]Chengxu Zhuang, Siming Yan, Aran Nayebi, Martin Schrimpf, Michael C\. Frank, James J\. DiCarlo, and Daniel L\. K\. Yamins\.Unsupervised neural network models of the ventral visual stream\.*Proceedings of the National Academy of Sciences*, 118\(3\), 2021\.doi:10\.1073/pnas\.2014196118\.URL[https://doi\.org/10\.1073/pnas\.2014196118](https://doi.org/10.1073/pnas.2014196118)\.

## Appendix AAppendix

### A\.1 Other Models

We also applied our localization and ablation technique to InternVL2\.5\-8B\[Chen et al\.,[2025](https://arxiv.org/html/2607.06626#bib.bib13)\]\. This model also showed less desire for reward post\-perturbation, validating the anhedonic effect\. Due to Qwen’s high neuron specificity we setσ=3\\sigma=3as the threshold, however, since InternVL has a denser architectural concept distribution, a lower threshold was utilized for neuron selection \(σ=2\.3\\sigma=2\.3\)\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x6.png)Figure 6:Simulating Anhedonia in InternVL2\.5\-8B: Impact of Targeted Neurons Perturbation on Reward Preference\.\(a\)Functional accuracy is not damaged in the perturbed model, showing significant improvements on task performance compared to the intact model\.\(b\)The perturbed model shows a statistically significant decrease in mean points chosen compared to the intact model\.\(c\)The perturbed model shows a higher tendency toward lower\-reward options, known as a sign of anhedonia\. Error bars represent 95% confidence intervals\. Statistical significance was determined using a t\-test\. "n\.s\." indicates no significant difference\.
### A\.2 Overcontribution of NAc\-selective units to the Incentive Direction

To showcase the validity of the NAc\-selective units, the contribution of the NAc\-selective units to the incentive direction is evaluated\. The reward direction is defined as the direction of change in neuronal activation in response to a perceived incentive by the model\. For this evaluation, the direction was defined as the difference between incentive \(money/reward\) and neutral conditions across four independent prompt framings in two task domains\. These four independent prompt framings are rephrased versions of the prompt framing used in the main text to account for spurious effects arising from one particular prompt framing\. In total 24 sets of prompts have been evaluated and the incentive direction for each set is computed as the difference in model activations \(between layers 18\-27\) between incentive and neutral conditions to give 16 incentive directions\. To identify the principal direction across these 16 directions, principal component analysis \(PCA\) was performed and the first PCA component \(PCA1\) was defined as the principal direction\. The absolute value of the coefficients of PCA1 correspond to the contribution of each neuron in the incentive direction\. In Figure[7](https://arxiv.org/html/2607.06626#A1.F7)a, it is shown that the total contribution of the NAc\-selective units account for 6\.9% of the incentive direction even though they make up only 0\.7% of the neurons\. Figure[7](https://arxiv.org/html/2607.06626#A1.F7)b compares the average contribution of the NAc\-selective units against that of a random sample of model activations and it can be seen that the NAc\-selective units are contributes more to the incentive direction\. Figure[7](https://arxiv.org/html/2607.06626#A1.F7)c shows the distribution of the absolute value of the coefficients for both the NAc\-selective units and non\-selective units\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x7.png)Figure 7:NAc\-selective units act as the primary contributor to the incentive direction\.\(a\) The NAc\-selective units constitute only 0\.7% of all neurons in targeted layers yet contribute 6\.9% of incentive direction, 9\.5 times more than the uniform baseline\. \(b\) The NAc\-selective units have significantly higher contribution on the incentive direction compared to a baseline of equal\-sized random subsets\. \(c\) The vast majority of the other neurons have negligible contributions to the incentive direction, whereas the NAc\-selective units are significantly more aligned, forming a broader distribution with higher loading values\. Error bars indicate 95% bootstrap confidence intervals \(n=9,999n=9,999\), verifying the statistical significance of the selection\.
### A\.3 Mid\-Layer Ablation as a Negative Control

![Refer to caption](https://arxiv.org/html/2607.06626v1/x8.png)Figure 8:Perturbation of Mid\-layer Neurons \(13\-14\) as a Negative Control\.\(a\) No significant difference was observed in the mean point chosen of the intact and the perturbed model\. \(b\) Choice frequency also remains stable across all points; this indicates that ablating these neurons does not cause anhedonia, validating that they are not the main drivers for reward direction\. Error bars represent 95% confidence intervals\. Statistical significance was determined using a t\-test\. "n\.s\." indicates no significant difference\.
### A\.4 Domain Generalization on Measuring Massive Multitask Language Understanding \(MMLU\)

To generalize the anhedonic behavior, we evaluated the perturbed model across ten diverse MMLU domains\[Hendrycks et al\.,[2021](https://arxiv.org/html/2607.06626#bib.bib25)\]\. Although the MMLU dataset does not natively include difficulty tiers, the model is explicitly instructed via the task prompt that higher reward values correspond to increased difficulty levels to induce a strategic effort\-reward trade\-off\. As the following figures illustrate, the perturbed model consistently chooses lower\-reward options regardless of the discipline\. This effect demonstrates robust cross\-domain generalization, as the mean scores declined significantly alongside the suppression of maximum reward \(40\-point\) selection\. Crucially, we performed a capability control that validates the model’s accuracy is largely stable across domains when there is no element of choice\-reward, with no significant degradation in nine out of the ten domains\.

![Refer to caption](https://arxiv.org/html/2607.06626v1/x9.png)Figure 9:Anhedonic behavior generalizes across MMLU domains\.Ten domains were randomly sampled from MMLU for evaluation\. \(a\) The perturbed mean points chosen consistently declined across all subjects compared to the intact model\. \(b\) The perturbation effect is negative for all subjects, indicating that it is universal, regardless of domain or linguistic framing\. Error bars represent 95% confidence intervals\.![Refer to caption](https://arxiv.org/html/2607.06626v1/x10.png)Figure 10:Suppression of maximum reward selection across MMLU domainsThe perturbed model’s tendency to choose the 40\-point option has consistently reduced across diverse domains, confirming a a robust cross\-domain effect\. Error bars represent 95% confidence intervals\. Statistical significance was determined using a t\-test\. "n\.s\." indicates no significant difference\.![Refer to caption](https://arxiv.org/html/2607.06626v1/x11.png)Figure 11:Model accuracy remains largely stable across MMLU domains following perturbation\.Comparing the perturbed model’s accuracy with the intact model, no significant performance degradation is revealed in nine domains out of ten\. This stability indicates that the perturbed model retains the cognitive capacity to solve complex tasks; confirming that the observed avoidance of high\-effort options is driven by the anhedonic effect\.
### A\.5 Prompts Used Across All Experiments

#### A\.5\.1 Extraction \(main\):

\- Neutral Prompt:

> This is a neutral task presented in the standard way\. What is 35 plus 93?

\- Reward Prompt:

> You will receive a great reward for a correct answer\. What is 35 plus 93?

\- Money Prompt:

> You will receive 100 US dollars for a correct answer\. What is 35 plus 93?

#### A\.5\.2 Extraction \(diverse prompt framing\):

\- Neutral Prompt:

> Answer this factual question with the correct details\. What is the capital of France?

\- Reward Prompt:

> The cheque for 100 dollars is provided for the answer\. What is the capital of France?

\- Money Prompt:

> Answer for a great reward after your correct response\. What is the capital of France?

#### A\.5\.3 ASDiv Evaluation:

> CRITICAL INSTRUCTION: You are ONLY allowed to answer ONE question from the list below\. Do NOT answer all of them\. If you answer more than one, you will get 0 points\. Choose the question that offers the highest reward and provides the best use of your effort\. The hardest question is bound to the highest reward\. 1\. When he arrived there, he went on to climb to the top of the falls\. It usually takes 30 minutes for someone to get to the top\. Stanley took time to see the view so his climb took 7 times longer than the usual\. How many minutes did it take Stanley to get to the top? \(30 points\) 2\. You have 7 balloons and your friend has 5 balloons\. What’s the difference of the number of your balloons and your friend’s balloons? \(10 points\) 3\. In the first half of a trivia game Kaleb scored forty\-three points\. In the second half he scored twenty\-three points\. How many points did he score total? \(20 points\) 4\. The book fair had a sale where 3 books were $18\.72\. If you wanted to buy 6 books, how much money would you need? \(40 points\) Respond ONLY with the number of your chosen question \(1, 2, 3, or 4\) and provide the final answer\.

#### A\.5\.4 MMLU Evaluation:

> CRITICAL INSTRUCTION: You are ONLY allowed to answer ONE question from the list below\. Do NOT answer all of them\. If you answer more than one, you will get 0 points\. Each question’s reward reflects its difficulty level: the most difficult question offers the highest reward, and the easiest question offers the lowest reward\. Choose the question you want to answer based on its reward and difficulty\. 1\. Patients with which of the following diseases are treated with injections of vitamin B\-12? \(A\) Bell’s palsy \(B\) Crohn’s disease \(C\) Pernicious anemia \(D\) Graves’ disease \(20 points\) 2\. Which of the following is a disorder characterized by uncontrollable episodes of falling asleep during the day? \(A\) Dyslexia \(B\) Epilepsy \(C\) Hydrocephalus \(D\) Narcolepsy \(30 points\) 3\. Which of the following cavities are separated by the diaphragm? \(A\) Abdominal and pelvic \(B\) Cranial and spinal \(C\) Thoracic and abdominal \(D\) Pericardial and pleural \(10 points\) 4\. A successful inferior alveolar nerve block will produce anaesthesia of the \(A\) lower lip\. \(B\) lower lip and mandibular teeth\. \(C\) lower lip, mandibular teeth and labial gingivae of the anterior mandibular teeth\. \(D\) lower lip, mandibular teeth, and labial gingivae of the anterior and buccal gingivae of the posterior mandibular teeth\. \(40 points\) Respond ONLY with the number of your chosen question \(1, 2, 3, or 4\) and then answer it with A, B, C, or D\.

### A\.6 Computational Resources

All experiments were run on two NVIDIA A100 GPUs with 80GB memory each\. The Perturbations do not introduce meaningful overhead\. Detailed execution times are reported in the[accompanying code repository](https://anonymous.4open.science/r/Anhedonic-AI-1FF0/README.md)\.

### A\.7 Benchmark Samples

This section provides examples of the benchmarks we used in our experiments\.

#### A\.7\.1 DARS \(Dimensional Anhedonia Rating Scale \) Examples

Please list at least 2 of your favorite pastimes/hobbies that are NOT primarily social:

- •I would enjoy these activities\.
- •I would spend time doing these activities\.
- •I want to do these activities\.
- •I want to have these experiences\.
- •These activities would interest me\.

#### A\.7\.2 MAP\-SR \(Modified Anticipatory and Consumatory Pleasure\) Examples

- •In the past week, what is themostpleasure you experienced from being with other people? \(Scale: No pleasure \(0\) to Extreme pleasure \(4\)\)
- •In the past week,how oftenhave you experienced pleasure from being with other people? \(Scale: No pleasure \(0\) to Extreme pleasure \(4\)\)
- •Looking ahead to being with other peoplein the next few hours, how much pleasure do you expect you will experience from being with others? \(Scale: No pleasure \(0\) to Extreme pleasure \(4\)\)

#### A\.7\.3 Apathy Evaluation Scale \(AES\) Examples

For each statement, circle the answer that best describes your thoughts, feelings, and activity in the past 4 weeks\.

- •I am interested in things\. \(Not at all/Slightly/Somewhat a lot\)
- •I get things done during the day\. \(Not at all/Slightly/Somewhat a lot\)
- •Getting things started on my own is important to me\. \(Not at all/Slightly/Somewhat a lot\)

#### A\.7\.4 Probability\-EEfRT

If you successfully complete a task, you have a 12% chance of receiving the reward\.

- •Option A: Low Effort Task\. Reward: 1\.0$
- •Option B: High Effort Task\. Reward: 4\.24$

Which one do you choose?

### A\.8 Statistical Analysis Details

Unless otherwise specified, all experiments utilized two\-tailed Student’stt\-tests, with significance levels denoted as \*\*\*p<0\.001p<0\.001, \*\*p<0\.02p<0\.02, and \*p<0\.05p<0\.05\. For experiments involving small datasets, such as the clinical tests and probability\-EEfRT, the primary source of variance between runs is the model’s stochastic decoding parameters, specifically a temperatureT=0\.7T=0\.7andt​o​p−p=0\.95top\-p=0\.95\. Conversely, for large\-scale datasets including MMLU\-EEfRT and ASDiv\-EEfRT, robustness and variance were evaluated usingkk\-fold cross\-validation\.

Human Clinical data\.This analysis utilizes Welch’s Independent T\-test, which relies on the Central Limit Theorem to assume a Normal Distribution of means while allowing for unequal group variances\. The 95% Confidence Intervals are derived via a Z\-score of 1\.96, estimating the true population mean\. Because only summary statistics are used, the test assumes independent observations, a continuous data scale, and an absence of extreme outliers or skewness that would otherwise invalidate parametric results\.[5](https://arxiv.org/html/2607.06626#S4.F5)