Expected Value of Control (EVC) Theory – Amitai Shenhav, Matthew M. Botvinick, & Jonathan D. Cohen

The capacity to adaptively align thought and action with internal goals, especially when confronted with compelling distractions, habitual intrusions, or demanding environmental obstacles, stands as the crowning achievement of the primate central nervous system. For decades, cognitive psychology and behavioral neuroscience categorized these regulatory operations under the umbrella of executive function or cognitive control. These functions were traditionally operationalized through algorithmic models of selective attention, working memory maintenance, prepotent response inhibition, and task switching. However, classical models routinely suffered from a glaring conceptual division: they treated the mechanisms of control execution—the top-down biasing of sensorimotor pathways—as fundamentally distinct from the motivational, affective, and economic mechanisms that determine whether, when, and to what extent such control ought to be deployed. Executive systems were frequently reified as quasi-homuncular supervisory modules that selected appropriate strategies without a transparent, formal accounting of the computational costs inherent to mental exertion.

To dismantle this mechanistic divide, Amitai Shenhav, Matthew M. Botvinick, and Jonathan D. Cohen formulated the Expected Value of Control (EVC) theory in their seminal 2013 framework. Grounded in computational neuroscience, bounded rationality, and normative economic theory, EVC recasts cognitive control from an obligatory or purely reactive processing filter into an optimal resource allocation problem. Under this paradigm, allocating cognitive control is formalised as an internal economic decision executed by an optimizing supervisory agent. This system must continuously weigh the anticipated payoffs of engaging effortful, goal-directed computation against the intrinsic, subjective costs associated with mental labor. By synthesizing reinforcement learning, evidence accumulation dynamics, and biophysical circuit models, the EVC framework provides an analytically rigorous foundation for evaluating how biological organisms arbitrate between automated habits and deliberate cognitive interventions.

At the anatomical core of this theoretical unification lies the dorsal anterior cingulate cortex (dACC). Historically one of the most empirically contested structures in human neuroimaging, the dACC has been alternately characterized as a hub for autonomic arousal, nociceptive processing, negative affect, motor selection, error identification, and conflict monitoring. Shenhav, Botvinick, and Cohen resolved this functional heterogeneity by proposing that the dACC acts as the computational engine of the EVC calculus. Rather than merely detecting processing conflict or signaling visceral distress, the dACC continuously monitors task contingencies, environmental volatility, reward prospects, and internal energetic states to estimate the expected value of allocating specific control signals. In doing so, it specifies both the qualitative identity and quantitative intensity of downstream control pathways, bridging the gap between value-based decision-making and executive neuroanatomy.

1. Introduction to Expected Value of Control (EVC) Theory and Historical Context

1.1 The Conceptual Origins of Cognitive Control

The theoretical trajectory of cognitive control reflects a long-standing effort to reconcile mechanistic accounts of information processing with the flexible, goal-directed nature of human behavior. Early cognitive psychology relied heavily on symbolic processing metaphors inspired by classical von Neumann computing architectures. In these initial formulations, executive functions operated as centralized software routines running atop passive sensory buffers and motor registers. Mental operations were understood as deterministic strings of logical operations executed serially, leaving little room for dynamic resource constraints or graded neural tuning. While these symbolic architectures successfully captured high-level logical reasoning and rule-based problem solving, they proved incapable of capturing the biophysical limitations, processing bottlenecks, and emergent stochasticity inherent to biological neural networks.

As connectionist models and parallel distributed processing frameworks gained traction in the late twentieth century, the field transitioned toward understanding cognition as the emergent outcome of reciprocal activations within vast neural populations. This shift brought into sharp focus the foundational dichotomy between automatic and controlled cognitive operations. Automatic processes were operationalized as fast, parallel, effortless, and computationally cheap transformations driven by heavily reinforced synaptic weights within dedicated processing pathways. Conversely, controlled processes were conceptualized as slow, serial, effortful, and capacity-limited operations required whenever novel tasks, ambiguous stimuli, or overriding prepotent reflexes demanded active guidance.

The most influential mid-century attempt to formalize this dynamic was the Supervisory Attentional System (SAS) model advanced by Donald Norman and Tim Shallice. The Norman-Shallice architecture posited that routine behavioral selections are governed by “contention scheduling”—a decentralized, competitive inhibition mechanism wherein well-learned schemas compete for motor expression based solely on sensory triggers and lateral inhibitory dynamics. However, when environmental conditions rendered these automatic schemas inappropriate, ambiguous, or dangerous, the SAS intervened. This supervisory controller provided top-down, bias-modulating signals to boost the activation of goal-relevant schemas above their natural competitive thresholds.

Although the SAS framework offered an elegant psychological taxonomy for executive control, it remained computationally underspecified. It treated the supervisory agent essentially as a black box: an unmodeled homunculus that somehow recognized when to intervene and how much intervention was required, without a clear computational explanation of its internal decision rules or resource trade-offs. It became increasingly obvious that any modern neurocomputational account of executive function needed to explain not only how supervisory biasing signals modulate lower-level sensory streams, but also how the brain decides that a specific situation warrants the costly deployment of those signals in the first place.

1.2 The 2013 Landmark Framework by Shenhav, Botvinick, and Cohen

To overcome the explanatory limitations of classical executive models, Amitai Shenhav, Matthew M. Botvinick, and Jonathan D. Cohen introduced the Expected Value of Control (EVC) theory in their landmark 2013 paper published in Nature Reviews Neuroscience. The motivation driving this framework was the realization that executive function could not be understood in isolation from normative decision theory. While neuroeconomics had made major strides in detailing how animals and humans compute subjective value, evaluate risk, and select between discrete goods or primary rewards, these economic principles were rarely applied to the brain’s internal operations. Shenhav and colleagues recognized that allocating cognitive control is itself an economic choice—not an external choice between two physical commodities, but an internal choice regarding the allocation of limited neurocomputational resources.

The 2013 paper synthesized two previously divergent fields: normative economic decision models and biophysically informed neural network architectures. Drawing from microeconomic expected utility theory, EVC asserted that the brain assigns an explicit subjective valuation to different states of mental labor. When a task demands sustained attention or response inhibition, the executive control system treats the allocation of its own processing capacity as a costly investment. Therefore, the brain must evaluate whether the anticipated gains—such as monetary rewards, social approval, or avoiding physical danger—exceed the intrinsic costs associated with engaging the requisite cognitive apparatus. By embedding this optimization logic within a mathematically explicit architecture, the authors moved past descriptive flowcharts of executive function to deliver a predictive, parameterized model of control allocation.

The primary objectives of the foundational publication were threefold. First, it aimed to specify a formal mathematical objective function defining the optimal control state based on environmental inputs, internal states, and processing costs. Second, it sought to delineate the neurocomputational microcircuitry capable of executing these calculations, explicitly proposing the dorsal anterior cingulate cortex (dACC) as the core anatomical nexus for this optimization. Third, the framework aimed to unify a vast, contradictory corpus of empirical neuroimaging and electrophysiological data that had accumulated over two decades regarding the functional specialization of the medial prefrontal cortex. In doing so, Shenhav, Botvinick, and Cohen established an integrative theoretical baseline that continues to structure modern cognitive neuroscience.

1.3 Resolving Functional Heterogeneity in the Anterior Cingulate Cortex

Prior to the formulation of EVC theory, the cognitive neuroscience literature was embroiled in persistent debates regarding the primary computational role of the anterior cingulate cortex, particularly its dorsal and rostral divisions. Functional magnetic resonance imaging (fMRI) investigations routinely documented elevated blood-oxygen-level-dependent (BOLD) responses within the dACC across an extraordinarily broad, seemingly disjointed array of experimental conditions. One major paradigm, spearheaded by Matthew Botvinick, Cameron Carter, and Jonathan Cohen, argued for the conflict monitoring hypothesis. This view maintained that the dACC acts as an online computational sensor dedicated to detecting simultaneous, mutually incompatible activations within downstream motor and cognitive networks—such as the co-activation of color and word representations in the classic Stroop task—thereby signaling the need for regulatory intervention by the dorsolateral prefrontal cortex.

Simultaneously, an entirely separate cadre of affective and clinical neuroscientists documented robust dACC engagement during the subjective experience of physical pain, social exclusion, negative emotional valence, and autonomic arousal. These findings prompted alternative models positing that the cingulate was fundamentally an emotional or visceral monitoring hub, alerting the organism to homeostatic imbalances, physical threats, or physiological stress. Adding further complexity, motor physiologists and non-human primate electrophysiologists frequently observed dACC single-unit activity directly correlated with motor error signaling, the tracking of reward prediction errors, the evaluation of counterfactual outcomes, and decisions to disengage from an existing behavioral patch to explore alternative environments.

This empirical splintering produced deep theoretical confusion. Researchers struggled to determine whether the dACC was composed of highly segregated, modular sub-regions dedicated to fundamentally distinct sensory, motor, affective, and cognitive functions, or whether there was an underlying, unifying computational operation that accounted for this cross-modal involvement. Competing meta-analyses debated boundary definitions, while lesion studies revealed puzzling preservation of executive abilities alongside profound motivational deficits, such as akinetic mutism and apathy syndromes.

The EVC framework resolved this functional heterogeneity through computational parsimony. Shenhav, Botvinick, and Cohen argued that conflict detection, pain processing, autonomic regulation, negative affect, and error signaling are not disparate functions carried out by segregated modules within the dACC. Instead, they represent common inputs to, or facets of, a single unified evaluation: the estimation of the expected value of control. Processing conflict, physical pain, and negative affect all serve as diagnostic signals indicating that current task performance is sub-optimal, environmental demands are escalating, or established policies are failing. Each of these events signals an immediate drop in expected payoff and an urgent need to reconfigure mental effort. The dACC does not merely log these negative events; it uses them as critical parameters to dynamically solve the core optimization problem: calculating which control signals should be selected, how intensely they must be energized, and whether the anticipated benefits of continued mental labor justify its immediate cognitive cost.

2. Foundational Architecture of the EVC Model

2.1 The Core Optimization Dilemma in Executive Function

At the center of the EVC model lies an evolutionary and biophysical reality: cognitive control is an intrinsically limited and energetically expensive resource. While sensory cortices can process immense arrays of visual and auditory features concurrently in an automated, massively parallel fashion, the top-down coordination of goal-directed actions faces strict processing bottlenecks. The brain cannot simultaneously direct maximal selective attention toward multiple conflicting goals, nor can it maintain an arbitrary number of complex behavioral programs in active working memory without catastrophic cross-talk. Hence, executive function faces a fundamental optimization dilemma: how should an organism allocate its limited cognitive bandwidth to maximize positive outcomes while minimizing waste, interference, and computational burnout?

The EVC model formalizes this challenge as a resource-constrained economic decision problem. In standard economic theory, a firm with finite capital must distribute its investments across competing production lines to maximize return on investment (ROI). Analogously, the brain must distribute its limited capacity for top-down biasing across competing sensorimotor, mnemonic, and executive pathways. If the system allocates insufficient control, it risks committing performance errors, succumbing to distracting sensory cues, or failing to capitalize on lucrative environmental opportunities. However, if it chronically allocates excessive control, it incurs severe computational and energetic penalties, rapidly exhausting subjective stamina and preventing other valuable cognitive operations from utilizing those same neural circuits.

This optimization dilemma inherently involves multidimensional trade-offs among computational precision, processing speed, and cognitive expenditure. Increasing the intensity of a control signal can accelerate sensory evidence accumulation and suppress motor noise, thereby boosting task accuracy and shortening response latencies. Yet, this high-precision state demands aggressive energetic expenditure and limits cognitive flexibility, temporarily blinding the agent to alternative opportunities or unexpected environmental shifts. The fundamental task of the EVC architecture is to calculate the precise inflection point where the marginal gain in expected performance payoff matches the marginal increase in the intrinsic cost of mental effort.

2.2 Feedforward Inputs and Feedback Evaluation Loops

To compute the optimal allocation of control, the EVC architecture relies on a continuous stream of feedforward information describing the state of the internal and external world, coupled with dynamic feedback loops that track the consequences of ongoing motor and cognitive executions. The feedforward stream provides the controller with the necessary contextual variables required to forecast upcoming demands before action execution finishes. These predictive inputs include representations of sensory stimuli, explicit task instructions, contextual environmental contingencies, current motivational drives, and subjective estimates of reward magnitude and probability.

Simultaneously, the EVC system relies on robust feedback loops that monitor real-time behavioral performance, error generation, and the expenditure of effort. When a controlled action is launched, these feedback loops monitor the rate of evidence accumulation in downstream cortical columns, registering whether processing is unfolding cleanly or whether it is encountering unexpected computational friction, such as target-distractor interference or internal response conflict. If an error occurs, or if the computational process takes considerably more time and energy than initially forecasted, feedback pathways deliver error-prediction signals and updated cost estimates back to the supervisory controller. This continuous monitoring updates the system’s internal model of task difficulty and environmental volatility.

Crucially, the EVC framework is not merely a reactive, error-correcting servomechanism; it is fundamentally an anticipatory, predictive controller. By integrating predictive state transitions with historical reinforcement learning profiles, the supervisory architecture forecasts upcoming control demands before overt behavioral failures can occur. For instance, when presented with a task cue indicating that an upcoming trial will involve severe stimulus-response conflict (such as an incongruent Stroop cue), the feedforward system utilizes internal models of state transitions to proactively scale up control intensity. This proactive deployment mitigates computational bottlenecks in advance, demonstrating how the integration of feedforward forecasting and feedback monitoring sustains robust, adaptive behavior.

2.3 Structural Distinction Between Controller and Controlled Systems

A foundational tenet of the EVC architecture is the structural and computational separation between the system that specifies control and the systems that execute it. The EVC model formally delineates the architecture into two distinct components: the controller (the supervisory agent localized primarily within the dACC) and the controlled system (the diverse, downstream sensory, memorial, motor, and premotor cortical networks that execute task-specific computations). This distinction prevents conflating the valuation and selection of control policies with the direct biophysical realization of sensory filtering or motor output.

The controlled system consists of task-specific execution modules distributed throughout the neocortex. For example, during visual attention tasks, the controlled system includes the ventral visual processing stream, the fusiform face area, and the parahippocampal place area, which process high-level visual representations. In motor tasks, the controlled system encompasses the primary motor cortex, supplementary motor area, and basal ganglia loops responsible for driving muscle effectors. Left to themselves, these downstream pathways operate largely via automatic, associative, and parallel mechanisms driven by bottom-up sensory stimulation and overlearned synaptic associations. They possess the machinery to execute behavior, but they lack the global objective function required to flexibly reorganize their own operations in response to shifting abstract goals.

The controller, in contrast, sits atop this processing hierarchy. Its primary output does not consist of direct motor commands or sensory perceptions, but rather of graded, top-down control signals. These control signals act as modulatory biases that selectively shift the operational state of the controlled system. By targeting specific downstream neural ensembles, the controller alters their intrinsic signal-to-noise ratio, suppresses task-irrelevant processing pathways via lateral or feedforward inhibition, and accelerates the rate of information processing along task-relevant routes. This communication is inherently bidirectional: while descending control signals enforce task rules and bias competitive dynamics downstream, ascending feedback signals inform the controller of processing efficiency, computational conflicts, and outcome realizations, enabling real-time recalibration of the EVC objective function.

3. The Computational Formulation of the Expected Value of Control

3.1 Mathematical Derivation of the Objective Function

The mathematical formalization of EVC operationalizes the allocation of cognitive control through an explicit objective function derived from normative decision theory. The core proposition is that the supervisory controller selects a vector of control signals that maximizes the difference between the expected utility of the resulting performance outcomes and the intrinsic cost incurred by implementing those signals. Let $S$ represent the current environmental and internal state, and let $signal$ (or $\mathbf{u}$) represent a specific control configuration, defined by both its identity and intensity. The expected value of specifying a given control signal in state $S$ is expressed mathematically as:

$$\text{EVC}(signal, S) = \sum_{i} P(Outcome_i mid S, signal) \cdot U(Outcome_i) – \text{Cost}(signal)$$

In this equation, $Outcome_i$ represents a potential state of affairs following task execution, which includes factors such as correct task completion, behavioral errors, temporal delays, secondary costs, and primary or secondary reward receipts. The term $U(Outcome_i)$ designates the subjective utility of that specific outcome, which is shaped by the agent’s current homeostatic requirements, economic goals, and task instructions. The term $P(Outcome_i mid S, signal)$ denotes the conditional transition probability distribution: the likelihood that outcome $i$ will materialize given that the environment is in state $S$ and the controller deploys the specific control vector $signal$.

The final term, $\text{Cost}(signal)$, represents the intrinsic, subjective cost of applying that particular control configuration. This cost function is monotonically increasing with respect to the intensity of the control signal and may also scale with the structural complexity or number of simultaneous processing pathways targeted by the signal vector. The optimization problem solved by the supervisory system at any decision epoch is to identify the optimal control signal configuration, denoted $signal^*$, that satisfies the argmax criterion:

$$signal^* = arg\max_{signal} \left[ \sum_{i} P(Outcome_i mid S, signal) \cdot U(Outcome_i) – \text{Cost}(signal) \right]$$

Through this formulation, Shenhav and colleagues demonstrated that an agent should rationally increase cognitive effort if and only if the marginal increase in expected outcome utility—driven by improvements in the probability distribution toward favorable outcomes—exceeds the marginal cost imposed by the increased control intensity.

3.2 Parameterized Action Selection and Transition Probabilities

To link the abstract probability distribution $P(Outcome_i mid S, signal)$ to biophysical reality, the EVC framework integrates computational models of perceptual and motor decision-making, most notably the Drift-Diffusion Model (DDM) and related linear ballistic accumulator frameworks. In standard sequential sampling models, decisions are made by accumulating noisy sensory evidence over time until the cumulative evidence crosses a predetermined decision threshold. The efficiency of this accumulation process is governed by the drift rate ($v$), which reflects the quality and signal-to-noise ratio of the incoming information stream, while the distance between decision bounds ($a$) reflects response caution or speed-accuracy trade-offs.

Within the EVC framework, the application of a control signal explicitly parameterizes these accumulation dynamics. Specifically, increasing control intensity directed toward a task-relevant sensory modality boosts the drift rate ($v$) toward the correct decision boundary while suppressing the drift rate induced by task-irrelevant, distracting visual or auditory inputs. Mathematically, the drift rate can be formulated as a function of the control signal: $v = f(signal, S)$. Because analytical solutions of the drift-diffusion model provide exact mathematical formulations for response time distributions and error probabilities as functions of drift rate, the EVC model directly maps control intensity to task accuracy and latency metrics:

$$P(\text{Correct} mid S, signal) = \frac{1}{1 + \exp\left( -2 \cdot v(signal, S) \cdot \frac{a}{\sigma^2} \right)}$$

Here, $\sigma^2$ represents the diffusion variance (neural noise) within the processing network. As the control intensity rises, $v(signal, S)$ increases, shifting the probability mass toward accurate choices and compressing response latencies. However, the system must continuously evaluate these performance gains through Bayesian updating. Because the true state of the environment and the precise mapping between control intensity and drift rate are often ambiguous, the controller updates its beliefs about environmental contingencies ($S$) and outcome probabilities using historical observation trajectories, ensuring that control parameters adapt to changing environmental demands.

3.3 Algorithmic Policy Optimization and Argmax Resolution

Calculating the optimal control signal via the argmax operation presents a major computational hurdle. In real-world environments, the control space is not a simple binary switch; it is a high-dimensional, continuous parameter space spanning dozens of distinct sensory channels, motor effectors, working memory slots, and internal processing gains. Determining the globally optimal control vector $signal^*$ in real time through brute-force exhaustive search across this high-dimensional landscape would be computationally intractable, easily leading to combinatorial paralysis.

To address this complexity, the EVC model posits that the supervisory architecture uses algorithmic heuristics and continuous optimization techniques, such as iterative gradient ascent, local hill-climbing, or amortized inference routines. Instead of searching through every possible combination of control parameters, the controller leverages smooth, continuous cost-benefit surfaces. In these configurations, small adjustments in control intensity produce predictable, monotonic changes in expected drift rates and cost penalties. By computing the local gradient of the EVC function with respect to the control vector—calculating $\nabla_{signal} \text{EVC}(signal, S)$—the neural circuitry can rapidly shift its current control state in the direction of greatest marginal improvement.

Furthermore, when faced with urgent temporal deadlines that preclude continuous gradient optimization, the biological system can deploy analytical approximations or pre-compiled control policies. These “cached” control heuristics, shaped by prior reinforcement learning, allow the agent to map specific environmental cues directly to stereotyped control configurations. This approach bypasses full online optimization when processing speed is critical. By balancing deep iterative optimization against rapid heuristic approximations, the EVC architecture resolves the argmax operation within the strict temporal windows demanded by natural environments.

4. Neuroanatomy of EVC: The Role of the Dorsal Anterior Cingulate Cortex

4.1 Afferent Connectivity and Multimodal Information Ingestion

For the dorsal anterior cingulate cortex to calculate the expected value of control, it must possess an anatomical connectivity profile capable of integrating diverse streams of motivational, sensory, cognitive, and physiological information. Non-human primate tract-tracing and human diffusion tensor imaging (DTI) demonstrate that the dACC (Brodmann areas 24a/b/c and 32) occupies a privileged topological position within the macroscopic cortical hierarchy. It receives dense, direct afferent projections from the ventromedial prefrontal cortex (vmPFC) and the orbitofrontal cortex (OFC), structures heavily implicated in encoding subjective economic value, goal states, and expected monetary or primary payoffs.

In parallel with these value-based inputs, the dACC receives interoceptive and physiological data via robust monosynaptic inputs from the anterior insular cortex, the amygdala, and autonomic nuclei within the brainstem. These pathways convey real-time information regarding visceral arousal, physical exhaustion, systemic inflammation, nociception, and sympathetic tone. This physiological grounding ensures that the EVC calculus reflects not only external rewards, but also the internal biological state and energetic reserves of the physical organism.

Simultaneously, the dACC receives dense dopaminergic projections originating directly from the ventral tegmental area (VTA) and the substantia nigra pars compacta (SNc), alongside rich ascending noradrenergic projections from the locus coeruleus (LC). These ascending monoaminergic inputs signal reward prediction errors, environmental volatility, and global vigilance demands. Finally, reciprocal connections with the posterior parietal cortex and the dorsolateral prefrontal cortex supply the dACC with detailed information regarding ongoing working memory representations, attentional load, and active task rules. This convergence of value representations, interoceptive states, neuromodulatory signals, and cognitive task sets provides the dACC with the multimodal inputs required to run the EVC algorithm.

4.2 The Dorsal Anterior Cingulate Cortex as an EVC Nexus

Equipped with this multimodal input stream, the internal microcircuitry of the dACC serves as the computational engine for EVC optimization. Unlike primary sensory or motor cortices, which display highly organized, modular columnar arrangements, the cytoarchitecture of the agranular and dysgranular anterior cingulate cortex features extensive horizontal recurrent collateral networks, deep-layer pyramidal cell assemblies, and dense populations of inhibitory interneurons. This structural motif supports parallel, non-linear computations across overlapping populations of neurons.

Electrophysiological recordings in non-human primates confirm that individual dACC neurons do not operate as unimodal feature detectors. Instead, they exhibit rich, mixed selectivity. Single neurons within the dACC simultaneously encode multiplexed combinations of task difficulty, anticipated reward magnitude, effort requirements, error likelihood, and past reinforcement history. Rather than segregating cost tracking and reward anticipation into distinct anatomical zones, the microcircuits of the dACC integrate these vectors within shared population codes, directly implementing the subtraction of subjective costs from anticipated gross payoffs.

This computational mechanism clarifies the long-standing paradox of dACC activation profiles in neuroimaging. Phenomena that historically appeared distinct—such as the pain of physical injury, the emotional distress of social rejection, the cognitive strain of resolving high-conflict Stroop trials, and the detection of motor slips—all evoke robust hemodynamic responses within the dACC precisely because they represent variables that drastically alter the cost-benefit equation of behavioral control. When task conflict surges, or when pain signals bodily vulnerability, the net expected value of current policies plunges, triggering the dACC microcircuitry to calculate the necessity, identity, and magnitude of a control reconfiguration.

4.3 Efferent Control Pathways and Downstream Execution

Once the dACC resolves the EVC optimization problem and specifies the optimal control vector, it must communicate these directives to the rest of the brain to modulate performance. The dACC achieves this via an extensive, divergent network of efferent projections targeting downstream cognitive, neuromodulatory, and motor structures. The primary cognitive route involves massive reciprocal connections with the dorsolateral prefrontal cortex (dlPFC). The dlPFC maintains rule-based representations and provides top-down attentional biasing signals to primary and association sensory cortices. In the EVC architecture, the dACC acts as the strategic manager that decides *how much* control to invest and *where* to direct it, while the dlPFC functions as the executive arm that physically orchestrates those attentional filters and working memory buffers.

Beyond these cortical projections, the dACC exerts rapid, widespread influence across the brain via direct efferent projections to subcortical neuromodulatory centers, most notably the locus coeruleus-norepinephrine (LC-NE) system. Grounded in the adaptive gain theory developed by Gary Aston-Jones and Jonathan Cohen, descending projections from the dACC modulate the firing modes of LC neurons. When the dACC calculates high EVC for a focused, task-specific state, it drives the LC into a phasic bursting regime. This bursts norepinephrine across the cortical mantle, systematically increasing the gain of downstream pyramidal neurons, boosting the signal-to-noise ratio of sensory representations, and driving up the drift rate within the controlled system.

Finally, the dACC maintains direct, descending motor projections that bypass the prefrontal cortex entirely. Cingulate motor areas (CMAs), embedded deep within the banks of the cingulate sulcus, project directly to the supplementary motor area (SMA), the premotor cortex, the striatum, and the spinal cord. These projections allow the dACC to exert direct motor gating effects. By modulating these motor pathways, the dACC can selectively suppress premature motor releases, enforce behavioral pauses during states of high uncertainty, or directly invigorate motor actions when the expected value of immediate physical intervention is high.

5. Dissecting Control Signals: Identity Versus Intensity

5.1 Control Specification: The ‘Identity’ Dimension

A central theoretical innovation of EVC theory is the formal decomposition of the supervisory control signal into two distinct dimensions: identity and intensity. The identity dimension addresses the qualitative question of control allocation: precisely which internal cognitive pathways, sensory streams, mental operations, or motor effectors must be targeted by the supervisory system? In any given environmental context, an organism could mobilize an array of internal operations: it could direct visual attention toward spatial locations, prioritize color over semantic word features, retrieve specific episodes from long-term memory, maintain an active phonological loop, or prepare a specific motor effector while inhibiting another.

The specification of control identity requires the dACC to route top-down biasing signals to the precise downstream cortical regions responsible for mediating the task-relevant information flow. For example, in a task demanding spatial selective attention, the control signal’s identity directs modulation toward the frontal eye fields (FEF) and the intraparietal sulcus (IPS). If the task instead requires categorizing emotional facial expressions amidst complex visual scenes, the identity of the control signal directs biasing inputs toward the fusiform face area (FFA) and the amygdala, while actively damping responsiveness within the parahippocampal place area (PPA).

This selective routing reconfigures the functional architecture of the brain on a moment-to-moment basis. By deploying control signals with specific identities, the EVC system dynamically re-weights synaptic transmission across distributed cortical pathways, constructing temporary functional networks dedicated to executing the task at hand. The identity dimension ensures that mental effort is not discharged as diffuse, uncoordinated neural noise, but is delivered with high spatial and functional selectivity to the exact neural circuits required to solve current environmental challenges.

5.2 Control Allocation: The ‘Intensity’ Dimension

While identity specifies the structural target of control, the intensity dimension determines the quantitative investment of effort: how much top-down bias should be exerted upon that chosen pathway? Control intensity is a graded, continuous variable. In the language of neurobiology, intensity corresponds to the magnitude of the biasing current delivered to targeted cortical assemblies, the degree of localized inhibitory drive suppressing competing representations, and the amplitude of phasic neuromodulatory release recruited to alter local signal-to-noise ratios.

Electrophysiologically, variations in control intensity manifest as parametric modulations of neural gain. When the EVC system allocates high control intensity to a specific sensory channel, the firing rates of neurons tuned to the target features are elevated, while the baseline activity and burst capabilities of neurons tuned to distractor dimensions are suppressed. In sequential sampling models, this increase in intensity shows up as a steepening of the drift rate parameter ($v$), driving faster decision times and lower error rates. Under low-intensity control, the drift rate drops, resulting in slower, more error-prone, and noise-sensitive performance.

However, allocating high control intensity comes with clear physiological and psychological trade-offs. The subjective sensation of mental exertion scales directly with control intensity. The harder an individual concentrates, the more severe the internal sensation of strain, and the faster the onset of cognitive fatigue. Furthermore, running cortical networks at maximal gain drastically elevates metabolic glucose and oxygen consumption and saturates working memory buffers. Therefore, the EVC optimization algorithm avoids simply defaulting to maximal control intensity; instead, it scales intensity down to the minimum level necessary to achieve satisfactory task performance based on the value of the anticipated payoff.

5.3 Multi-Target Combinatorial Allocation Challenges

The complexity of executive control becomes acute when an agent attempts to pursue multiple goals simultaneously. This scenario exposes the multi-target combinatorial allocation problem. In multi-tasking environments, the supervisory controller must decide whether to route its limited control budget entirely to a single target, split intensity across multiple concurrent pathways, or rapidly alternate full intensity between competing channels over time. Because the space of possible control combinations grows exponentially with every additional task dimension, searching this space presents a major challenge for the central nervous system.

This allocation problem is further complicated by cross-talk and mutual interference between overlapping neural representations. When two tasks rely on shared representational substrates—such as two linguistic tasks utilizing the same left-hemisphere phonological networks, or two motor tasks competing for identical downstream effectors—allocating control intensity simultaneously to both streams inevitably produces catastrophic interference. Even when tasks use superficially distinct sensory inputs, the centralized, low-dimensional nature of the prefrontal control architecture creates a structural bottleneck that severely penalizes concurrent execution.

The EVC model formalizes these limitations by treating cross-talk as a sharp elevation in the intrinsic cost function. The cost of running two conflicting control signals simultaneously is far higher than the sum of their individual costs run in isolation: $\text{Cost}(signal_A + signal_B) gg \text{Cost}(signal_A) + \text{Cost}(signal_B)$. Consequently, the EVC system’s argmax calculation routinely favors policies of strict single-task prioritization or rapid serial switching over concurrent parallel allocation, unless the expected payoffs for both concurrent streams are high enough to offset the steep interference costs.

6. The Intrinsic Cost of Cognitive Effort and Mental Labor

6.1 The Phenomenological and Theoretical Nature of Mental Effort

A universal feature of human conscious experience is that sustained, intensive thinking feels difficult. Individuals universally report that concentrating on complex mathematical derivations, resisting strong emotional impulses, or maintaining vigilant focus on monotonous sensory streams entails a distinct, aversive sensation of mental strain. Historically, cognitive psychology struggled to incorporate this subjective phenomenology into its mechanistic models, frequently treating mental effort simply as an unmodeled descriptive label for high attentional load.

Over the past decade, however, behavioral economics and cognitive neuroscience have demonstrated that cognitive effort carries measurable subjective disutility. When given a free choice between two tasks that yield identical monetary rewards, human participants and non-human animals systematically avoid the task that demands higher cognitive control. This phenomenon is clearly demonstrated using paradigms such as the Demand Selection Task (DST) developed by Kirsten McGuire and Matthew Botvinick. In the DST, participants learn to associate specific visual cues with varying levels of cognitive demand (such as high-switch versus low-switch task blocks). Even in the complete absence of explicit instructions, participants develop a strong, implicit bias toward selecting the low-demand options.

This observation led EVC theory to formalize cognitive control as a scarce, intrinsically costly currency. But why does mental effort carry subjective disutility? Two competing theoretical frameworks attempt to explain this cost:

• The metabolic constraint account: Suggests that the brain, which consumes roughly 20% of resting bodily energy, faces localized metabolic bottlenecks (e.g., extracellular glucose depletion, lactate accumulation, or astrocytic glycogen exhaustion) during sustained prefrontal activation, producing an aversive signal to protect metabolic homeostasis.
• The representational/computational constraint account: Holds that the cost of effort is not a direct readout of peripheral metabolic depletion, but rather an internally computed opportunity cost. Because the executive system has limited capacity, dedicating control circuits to Task A prevents those exact same resources from being used for Task B. The subjective feeling of effort serves as an adaptive, aversive motivational signal, forcing the agent to weigh whether current pursuits justify the foregone value of all alternative actions.

6.2 Effort Discounting Functions and Mathematical Modeling

To parameterize the subjective cost of mental labor within the EVC objective function, researchers use mathematical effort discounting models. Similar to how delayed rewards undergo temporal discounting and probabilistic rewards undergo risk discounting, the subjective value of a reward is discounted as the required mental effort increases. If an agent is offered a gross reward $R$ for completing a cognitively demanding task requiring control intensity $C$, the net subjective value ($SV$) of that opportunity can be modeled using several distinct mathematical formulations.

A simple baseline model is the linear discounting function, where subjective value decreases in direct proportion to control intensity: $SV(R, C) = R – \alpha C$, where $\alpha$ represents an individual-specific effort sensitivity parameter. However, empirical behavioral data from cognitive effort discounting paradigms consistently show that linear functions fail to capture human behavior at high effort levels. Instead, effort discounting demonstrates marked non-linearity: small increments of effort are well-tolerated at low baselines, but cost penalties accelerate sharply as control demands approach an individual’s limits.

Consequently, researchers model effort costs using parabolic (quadratic) or hyperbolic discounting formulations. Under a parabolic discounting model, cost accelerates quadratically with control intensity:
$$SV(R, C) = R – \alpha C^2$$
This formulation mirrors standard physical work functions and reflects the steep computational penalties imposed when prefrontal circuits approach operational capacity. Alternatively, hyperbolic discounting can be formulated as:
$$SV(R, C) = \frac{R}{1 + \alpha C^2}$$
Here, the marginal cost of effort increases rapidly at high intensities. Within the EVC framework, this non-linear acceleration in the cost term ensures that the supervisory system will only allocate extreme control intensities when the prospective reward is exceptionally large, preventing the system from over-committing resources to marginal prospects.

6.3 Inter-Individual Variability in Effort Sensitivity

The parameter governing the steepness of the effort cost function—the effort sensitivity coefficient $\alpha$—is not a universal biological constant. It exhibits wide, stable variation across the human population. Psychometrically, this variability maps onto trait-level constructs such as the Need for Cognition (NFC), an inventory designed to quantify an individual’s intrinsic tendency to engage in and enjoy effortful cognitive endeavors. Individuals scoring high in Need for Cognition display a flatter effort discounting curve; they require significantly lower monetary compensation to undertake complex mental tasks, finding the computational labor intrinsically rewarding or less subjectively taxing.

Neurochemically, these inter-individual variations in effort sensitivity track differences in baseline monoaminergic tone, particularly within striatal and prefrontal dopaminergic circuits. Positron emission tomography (PET) imaging utilizing radiotracers such as [11C]-raclopride and [18F]-DOPA indicates that individuals with higher baseline dopamine synthesis capacity in the striatum display lower subjective effort costs. They are substantially more willing to expend cognitive effort for incremental monetary payoffs than individuals with reduced striatal dopamine synthesis capacity.

Furthermore, an individual’s effort cost function is not static; it is modulated by transient physiological and environmental states. Acute psychosocial stress, systemic inflammation, circadian nadirs, and sustained cognitive exertion (producing subjective mental fatigue) all induce an acute upward shift in the effort sensitivity parameter $\alpha$. Under these strained states, the subjective cost of deploying even moderate control intensities escalates, causing the EVC optimization equation to tilt away from demanding, goal-directed interventions toward automated, habitual, or exploratory behavioral policies.

7. Integration of Reward and Effort: Reinforcement Learning Dynamics

7.1 Dopaminergic Gating and Valuation Dynamics

The continuous evaluation and calibration of the EVC objective function are deeply intertwined with ascending monoaminergic neuromodulatory systems, most notably ascending dopaminergic pathways. Dopamine plays a dual role in cognitive neuroscience: it signals reward prediction errors (RPEs) to drive reinforcement learning across the striatum and cortex, and it acts as an activator that invigorates physical and cognitive actions by scaling expected net utility against required physical or mental expenditure.

The mesolimbic dopamine system (originating in the VTA and projecting to the ventral striatum and nucleus accumbens) and the mesocortical dopamine system (projecting directly to the dACC and dlPFC) play distinct, complementary roles in setting the EVC. While the ventral striatum uses dopaminergic bursts to register anticipated reward values, the dACC synthesizes these reward expectations with incoming cost assessments. Dopaminergic transmission within the dACC directly modulates the signal-to-noise ratio of local pyramidal cells through D1 and D2 receptor cascades, altering the threshold at which a control policy is deemed worth executing.

Pharmacological challenges in both humans and animal models consistently confirm this dopaminergic involvement. Administering dopamine reuptake inhibitors or receptor agonists (such as methylphenidate, modafinil, or amphetamine salts) systematically increases an agent’s willingness to select high-effort cognitive tasks over easy, low-reward alternatives. Within the formal mathematics of the EVC model, these dopaminergic interventions can be formalized either as scaling up the subjective utility of the expected reward ($U(Outcome)$) or as directly attenuating the steepness of the cost penalty ($\text{Cost}(signal)$). By mitigating the perceived burden of mental exertion, dopaminergic signaling gates the entry of supervisory control signals into downstream execution pathways.

7.2 Model-Based versus Model-Free Computations in EVC

A foundational distinction in modern reinforcement learning is the dichotomy between model-based and model-free decision-making. Model-free algorithms evaluate actions based entirely on cached, historic values reinforced through trial-and-error experience. They are computationally cheap and fast, but inflexible when environmental contingencies shift unexpectedly. Conversely, model-based algorithms utilize an internal cognitive map of the environment, simulating branching trees of future state transitions and prospective consequences. This approach affords high behavioral flexibility at the expense of substantial computational effort.

The EVC framework captures this dynamic by demonstrating that computing the optimal control allocation itself balances model-based and model-free mechanisms. Calculating full EVC via exhaustive prospective search over all possible outcome states is an inherently model-based operation:
$$\sum_{i} P(Outcome_i mid S, signal) \cdot U(Outcome_i)$$
This prospective computation requires the dACC to maintain an internal representation of the task’s state space, simulate the downstream consequences of various control intensities, and sum the resulting utilities. This meta-level model-based computation is itself cognitively demanding.

Consequently, the brain deploys model-free heuristics to regulate its executive control allocations. When a particular task context is encountered repeatedly, the supervisory system ceases running complex forward simulations. Instead, it relies on cached, model-free estimates of the control intensity that proved successful in that context previously. The arbitration between model-based control optimization and model-free heuristic deployment is governed by an economic trade-off: when the stakes are low or the environment is highly familiar, the system relies on computationally cheap, model-free control policies. However, when environmental volatility surges, stakes escalate, or novel contingencies emerge, the system invests in expensive model-based simulations to precisely calibrate its control signals.

7.3 The Learned Value of Control (LVC) Framework

Building directly upon the foundation of EVC, contemporary cognitive neuroscience has expanded to incorporate the Learned Value of Control (LVC) framework, developed extensively by researchers like Wilbertz, Egner, and Shenhav. The central insight of LVC is that the willingness to exert cognitive control is not an invariant trait, nor is it calculated entirely de novo at every isolated decision point. Instead, the control system displays meta-learning: the efficiency, efficacy, and perceived value of allocating control are systematically calibrated by an individual’s cumulative reinforcement history.

If an agent repeatedly discovers that deploying high control intensity in a specific environmental context consistently yields large, dependable rewards, the system updates its internal value priors. Plasticity within dACC-frontostriatal networks lowers the cost penalty associated with that specific control state, transforming what was once an agonizingly effortful mental exertion into a smoothly mobilized, highly reinforced cognitive routine. Conversely, if an agent repeatedly experiences environments where high mental exertion fails to improve performance outcomes—or where rewards are dispensed erratically regardless of effort—the LVC calculus degrades. The system learns that cognitive control has low efficacy, leading to learned helplessness, task disengagement, and chronic under-allocation of mental effort.

This learned plasticity in control allocation provides a mechanistic account for the acquisition of complex cognitive expertise. Experts in cognitively demanding domains (such as competitive chess players, software engineers, or surgeons) do not necessarily possess an anomalous energetic reservoir that protects them from mental exhaustion. Rather, through extensive reinforcement histories, their frontoparietal architectures have updated the learned value of control for domain-specific tasks. As a result, the dACC routinely forecasts high positive returns for sustaining focused control intensities, systematically suppressing the subjective disutility of domain-specific mental labor.

8. Empirical and Neuroimaging Paradigms Validating EVC Theory

8.1 Interference and Conflict Paradigms

The initial empirical validations of EVC theory leveraged classic cognitive psychology interference paradigms, most notably the Stroop task, the Eriksen Flanker task, and the Simon spatial conflict task. In the Stroop task, participants must name the ink color of a printed word while suppressing the automatic reading of the word itself. In incongruent trials (e.g., the word “RED” printed in blue ink), the prepotent, automated reading schema conflicts directly with the task-relevant color-naming schema, requiring top-down control to bias processing toward the weaker color representation.

The EVC framework made critical, novel predictions about these paradigms that moved beyond classic conflict monitoring models. Because EVC operationalizes control as an economic trade-off, it predicts that manipulating prospective rewards should systematically alter how the brain handles cognitive conflict. In a series of experiments manipulating incentive sizes on a trial-by-trial basis, researchers confirmed that when high monetary rewards are promised for fast and accurate performance on upcoming incongruent trials, participants show dramatic reductions in Stroop and Flanker interference effects. The dACC scales up proactive control intensity in response to high-reward cues, steepening the drift rate for the task-relevant feature and resolving conflict significantly faster.

Furthermore, EVC provides a quantitative explanation for the Gratton effect (conflict adaptation)—the well-documented empirical phenomenon where the interference effect is significantly smaller on a trial immediately preceded by a high-conflict trial. While older models treated this adaptation purely as a reactive, homeostatic response to detected conflict, EVC demonstrates that the preceding conflict trial delivers an empirical update to the dACC’s estimate of environmental difficulty. The system updates its transition probabilities, calculates an elevated expected value for maintaining high control intensity into the next trial, and maintains top-down biasing on downstream sensory cortices to minimize future performance slips.

8.2 Cognitive Effort Discounting Tasks (COG-ED)

Direct empirical verification that mental effort operates as an economic cost within the EVC framework came with the development of the Cognitive Effort Discounting (COG-ED) paradigm by Westbrook, Kester, and Braver. The COG-ED paradigm provides an objective, psychophysically titrated metric for quantifying the subjective cost of cognitive effort, isolating it completely from physical fatigue or motor expenditure.

In a standard COG-ED paradigm, participants first complete multiple levels of a demanding cognitive task, such as the $N$-back working memory task (ranging from the simple 1-back to the highly demanding 4-back) or complex task-switching batteries. Once participants have experienced the subjective effort demanded by each level, they are placed in an economic titration choice phase. Participants are repeatedly asked to choose between a low-demand baseline task (e.g., the 1-back) paired with a modest monetary reward, and a high-demand task (e.g., the 3-back or 4-back) paired with a larger monetary reward (e.g., $1.00 for the 1-back versus$5.00 for the 4-back).

By parametrically titrating the monetary offer paired with the low-effort option across successive trials, the paradigm determines an explicit indifference point. This indifference point represents the precise monetary amount at which the participant views the low-effort and high-effort options as having identical subjective value. The difference between the objective monetary offer for the high-effort task and the titrated indifference point directly quantifies the subjective cost of the mental labor required by that task level. Empirical results across hundreds of human subjects confirm that cognitive effort discounting curves are steep, reliable over time, and highly sensitive to parametric increases in working memory load ($N$-back levels), confirming the foundational EVC premise that cognitive control carries an intrinsic, quantifiable cost penalty.

8.3 Functional Neuroimaging and Electrophysiological Evidence

Neuroimaging and human intracranial electrophysiology have generated strong empirical support for the specific computational architecture proposed by Shenhav and colleagues. Functional magnetic resonance imaging (fMRI) studies explicitly testing the EVC model demonstrate that BOLD activation in the dACC does not merely scale with raw conflict or error generation; instead, it parametrically tracks the net expected value of control across trials. When task difficulty and prospective reward are manipulated independently, dACC activation reflects the integrated combination of these vectors—scaling positively with prospective reward when control is necessary to achieve it, and scaling with effort costs when investments must be evaluated.

In the domain of human electrophysiology, two classic event-related potential (ERP) components have been reinterpreted through the lens of EVC: the Error-Related Negativity (ERN) and the Feedback-Related Negativity (FRN). The ERN is a sharp, frontocentrally distributed negative deflection occurring within 50–100 milliseconds following the execution of an erroneous motor response. Previously viewed as a dedicated “error detection” signal, the EVC framework reframes the ERN as a rapid, negative utility prediction error calculated by the dACC. The ERN reflects an immediate downward update in the expected payoff of the current control policy, which initiates an emergency recalibration of downstream control intensity to prevent cascading behavioral failures.

Finally, direct intracranial electrophysiological recordings in surgical patients and non-human primates have provided the ultimate validation of EVC’s neural predictions. Studies recording from single units within the human and macaque dACC reveal individual neurons displaying multiplexed tuning: the same single neuron tracks the anticipated reward magnitude of the upcoming trial, the computational difficulty of the required rule, and the historical error rate. These single-unit confirmations definitively demonstrate that the dACC houses the integrated cellular assemblies necessary to compute the multi-parameter EVC objective function.

9. Theoretical Distinctions: EVC Versus Conflict Monitoring and Resource Models

9.1 From Conflict Detection to Cost-Benefit Valuation

To fully appreciate the theoretical contribution of EVC, it is necessary to contrast it with its direct intellectual predecessor: the classic conflict monitoring hypothesis formulated by Botvinick, Braver, Barch, Carter, and Cohen in 2001. The conflict monitoring model represented a breakthrough in cognitive science by demonstrating how a simple computational mechanism—calculating the energy hop or dot product of mutually competing motor representations—could provide an automated trigger for adjustments in cognitive control without requiring a conscious homunculus. In that early model, the dACC was operationalized purely as an informational conflict sensor: when computational energy spiked between incompatible response channels, the dACC detected this interference and signaled the dlPFC to increase its attentional bias on task-relevant features.

Despite its mathematical elegance, the 2001 conflict monitoring model suffered from major theoretical limitations. Most critically, it was completely value-blind. In that architecture, the dACC had no mechanism to evaluate *why* it should resolve conflict, whether resolving that conflict was economically worth the effort, or what outcomes were at stake. The system would invest identical control intensity to resolve conflict in a trivial experimental task with zero reward as it would in a life-or-death survival scenario, provided the raw interference vectors were mathematically equivalent. Furthermore, the classic model struggled to explain why the dACC showed massive activation in response to primary pain, monetary losses, and social stress in the complete absence of motor response competition.

The 2013 EVC theory resolved these limitations by subsuming conflict monitoring within a broader normative economic valuation framework. In EVC, conflict is no longer the sole variable that triggers dACC activation. Instead, computational conflict is reframed as an important piece of information indicating that the current state is computationally inefficient, error-prone, and demanding of extra resources. Conflict detection is transformed from an isolated, closed-loop reflex into an input parameter feeding the overarching EVC calculus. By synthesizing this information with prospective payoffs, energetic costs, and transition probabilities, EVC provides an account of *proactive*, value-guided control that purely reactive conflict monitoring models could not accommodate.

9.2 EVC versus Ego Depletion and Classical Capacity Theories

The EVC model also directly challenges and replaces classical “limited resource” theories of willpower, most notably the once-popular “ego depletion” or glucose depletion model advanced by Roy Baumeister and colleagues. The ego depletion framework posited that executive control and self-regulation rely on a finite physical reservoir of energetic fuel, analogous to a muscle that tires after physical exertion. Early proponents of this theory asserted that engaging in self-control physically depletes circulating blood glucose levels in the brain, leaving the individual physiologically incapable of exerting further self-regulation until those metabolic stores are replenished via carbohydrate ingestion or rest.

The glucose depletion model has faced severe empirical and theoretical challenges. Methodologically, large-scale multi-laboratory replication initiatives failed to replicate the basic behavioral ego depletion effect. Biophysically, neuroimaging and metabolic physiology proved that executing high-demand cognitive tasks causes negligible fluctuations in global brain glucose consumption—the brain consumes nearly constant amounts of energy regardless of whether an individual is daydreaming or solving advanced differential equations. The drop in circulating glucose required by the ego depletion hypothesis would induce severe neurological emergencies, not merely occasional lapses in self-discipline.

The EVC framework offers a biologically viable, computational alternative to the ego depletion hypothesis. Under EVC, the drop in cognitive performance observed after prolonged mental exertion is not caused by the physical running-out of an energetic fuel tank. Rather, it reflects an adaptive, motivational recalibration. As an agent sustains cognitive effort over time, the intrinsic cost term ($\text{Cost}(signal)$) in the EVC equation accumulates. The brain computes this escalating cost as a protective mechanism: sustained focus on a single task means the agent is chronically incurring severe opportunity costs, ignoring other potentially vital survival needs and environmental options. The sensation of mental fatigue is not a mechanical failure of the hardware, but a functional economic signal generated by the EVC system to discourage continuous, narrow hyper-focus, encouraging the agent to disengage and explore alternative behaviors.

9.3 Comparison with Foraging and Multiple-Demand Network Frameworks

To contextualize EVC within broader neuroscience literature, it is essential to examine its relationship to two competing contemporary frameworks: behavioral ecology foraging models and John Duncan’s Multiple-Demand (MD) network theory. In behavioral ecology, the Marginal Value Theorem (MVT) formalizes how an animal decides to stay in its current food patch versus abandoning it to forage for a new one. Researchers led by Nils Kolling and Matthew Rushworth applied this logic to the brain, proposing that the dACC functions primarily as a “foraging engine” that tracks the value of exploring alternative behavioral options, while the vmPFC tracks the value of exploiting the current choice.

Shenhav, Cohen, and Botvinick engaged in a vigorous theoretical debate with the foraging camp, successfully demonstrating that the EVC model naturally accounts for foraging behaviors without requiring an ad-hoc, specialized foraging module. In the EVC framework, the decision to abandon an existing task patch occurs whenever the expected value of continuing control within that patch drops below the expected value of engaging alternative behavioral policies. The metrics that drive patch departure in foraging paradigms—diminishing returns, rising effort costs, and escalating search times—are the exact computational variables that enter the EVC objective function. Thus, foraging is simply a special case of general control allocation.

Concurrently, John Duncan formulated the Multiple-Demand (MD) network hypothesis, which posits that a widely distributed set of frontoparietal regions—encompassing the dACC, inferior frontal sulcus, anterior insula, and intraparietal sulcus—operates as a unified, domain-general processing engine. The MD framework emphasizes functional equivalence across these regions, arguing that they collaborate indiscriminately to construct structured, step-by-step cognitive programs regardless of the task domain.

While the EVC theory agrees with Duncan that cognitive control is inherently domain-general, it rejects the idea of computational homogeneity across the network. EVC establishes a clear division of labor between the nodes of the Multiple-Demand system. The dACC does not duplicate the representational maintenance functions carried out by the lateral prefrontal cortex or the spatial transformations executed by the parietal cortex. Instead, EVC positions the dACC as the specialized *evaluative optimizer* nested within this broader frontoparietal architecture. The dACC evaluates costs and expected payoffs to determine *which* programs the rest of the Multiple-Demand network should execute and *how intensely* they should run, providing an analytical hierarchy absent from purely distributed accounts.

10. Clinical Implications: Neuropsychiatric Disorders through the EVC Lens

10.1 Major Depressive Disorder and Apathy Syndromes

The computational framework provided by EVC provides valuable insights into the pathophysiology of neuropsychiatric conditions characterized by motivational collapse and cognitive deficits. Foremost among these is Major Depressive Disorder (MDD). Historically, the cognitive impairments in depression—such as working memory deficits, indecisiveness, and executive dysfunction—were viewed as secondary byproducts of negative emotional rumination. However, computational psychiatry now reframes core depressive symptoms, particularly anhedonia and psychomotor retardation, as structural distortions within the EVC objective function.

In patients suffering from MDD, functional neuroimaging consistently reveals blunted hemodynamic responses within the ventromedial-dACC circuit during the anticipation of potential rewards, paired with exaggerated neural responses to cost and effort demands. Within the mathematical architecture of EVC, this pathology corresponds to a dual distortion: the subjective payoff of success ($U(Outcome)$) is chronically down-weighted, while the effort sensitivity parameter ($\alpha$) is pathologically elevated. As a consequence, even trivial everyday tasks requiring executive control—such as paying bills, completing chores, or initiating social contact—produce a negative net EVC:
$$\text{EVC}(signal, S) < 0$$
When the brain’s supervisory controller calculates that the cost of mobilizing mental effort reliably exceeds its anticipated payoff, the system rationally decides not to allocate control. Psychomotor retardation, executive disengagement, and severe apathy are thus the direct, computationally predictable outcomes of an EVC equation that chronically values the mobilization of effort as an economically irrational endeavor.

10.2 Attention-Deficit/Hyperactivity Disorder (ADHD)

Attention-Deficit/Hyperactivity Disorder (ADHD) provides another compelling example of control misallocation that can be understood through the EVC framework. Individuals diagnosed with ADHD struggle to sustain selective attention and cognitive effort during continuous, low-incentive tasks (such as academic study or administrative work), yet frequently demonstrate hyper-focused attention for prolonged intervals during high-incentive, immediately reinforcing activities (such as dynamic video games or emergency scenarios). This striking behavioral dissociation challenges the notion that ADHD is a simple, structural inability to generate cognitive control.

The EVC model explains this paradox by demonstrating how ADHD alters the temporal discounting of expected payoffs within the control allocation loop. Individuals with ADHD exhibit unusually steep, hyperbolic temporal discounting functions. In environments where the rewards for deploying cognitive effort are temporally remote—such as studying for an exam occurring weeks later—the discounted subjective value of that outcome collapses to near zero. Because the anticipated payoff is heavily discounted, the net EVC for mobilizing sustained attentional control becomes negative, leading to immediate task abandonment and mind-wandering.

Furthermore, erratic phasic and tonic dopaminergic signaling across frontostriatal circuits destabilizes the intensity dimension of the control signal, producing massive trial-by-trial variability in drift rates and reaction times. This computational framing explains the therapeutic mechanism of psychostimulant medications, such as methylphenidate and mixed amphetamine salts. By blocking dopamine and norepinephrine reuptake within frontostriatal pathways, these agents boost the basal signal-to-noise ratio within the dACC. This pharmacological intervention amplifies the perceived valuation of delayed goals and scales down the subjective burden of effort costs, restoring positive EVC calculations and enabling the stable, proactive mobilization of cognitive control.

10.3 Schizophrenia and Negative Symptomatology

In schizophrenia, negative symptoms—most notably avolition, asociality, and alogia—represent some of the most disabling and pharmacologically intractable aspects of the illness. While positive symptoms (such as hallucinations and delusions) are tied to aberrant striatal dopamine release, negative symptoms are increasingly recognized as an impairment in the representation of value-guided goal states needed to direct cognitive actions.

Viewed through the EVC framework, avolition in schizophrenia stems from an inability to construct and maintain stable, model-based representations of expected payoffs within the prefrontal cortex. Working memory impairments—driven by NMDA receptor hypofunction and dysregulated GABAergic interneuron firing within the dlPFC—degrade the integrity of the state representation ($S$). If the internal model of the current and future environmental states is corrupted, the supervisory controller cannot accurately estimate the state transition probabilities:
$$P(Outcome_i mid S, signal)$$
The controller cannot reliably forecast that deploying mental effort will produce successful task completion.

Consequently, the connection between effort expenditure and reward attainment is broken. Behavioral testing in patients with schizophrenia demonstrates a clear dissociation: while their hedonic experience of rewards in the moment remains largely intact (normal consummatory pleasure), their willingness to exert physical or cognitive effort to *obtain* those exact same rewards is severely degraded. The EVC calculus fails not because rewards lack hedonic value, but because the cognitive machinery required to simulate the value of control investments is damaged, leaving the individual trapped in an avolitional state.

10.4 Obsessive-Compulsive Disorder and Addiction

At the opposite end of the spectrum lies Obsessive-Compulsive Disorder (OCD), a condition characterized by the pathological *over-allocation* of cognitive control. Patients with OCD experience intrusive, distressing thoughts (obsessions) that trigger rigid, repetitive cognitive or behavioral routines (compulsions). Neuroimaging consistently documents chronic hyperactivity within the cortico-striatal-thalamo-cortical (CSTC) loops, particularly encompassing the dACC, caudate nucleus, and OFC.

Through the lens of EVC, OCD represents a severe, hyper-inflated estimation of the threat-avoidance value of control allocation. In the minds of individuals with OCD, failing to mobilize mental effort is forecasted to result in catastrophic, high-magnitude negative utility. The dACC microcircuitry calculates an astronomical expected value for intervening with repetitive checking, counting, or neutralizing rituals. The system becomes hyper-vigilant: the threshold for triggering supervisory control drops excessively, and the perceived cost of executing these mental rituals is ignored because the anticipated cost of non-intervention appears catastrophic.

Conversely, substance use disorders and behavioral addictions reflect a progressive, structural disconnection between automatic habit systems and executive oversight. Chronic drug exposure alters neuroplasticity within the striatum, shifting behavioral control from goal-directed ventral striatal circuits to automated, dorsolateral striatal habit networks. Concurrently, it induces severe structural and functional hypoactivity within the dACC and prefrontal cortex. Under the EVC model, addiction produces a state where compulsive reward pursuit becomes insulated from the supervisory controller. The automatic processing pathways of the controlled system operate uncontrollably, while the damaged EVC hub lacks the signal intensity required to override these deeply entrenched, automatic behavioral patterns.

11. Contemporary Extensions, Criticisms, and Algorithmic Refinements

11.1 Hierarchical EVC Frameworks

Since the publication of the original 2013 paper, research has substantially expanded the architectural sophistication of EVC theory. A major development is the formalization of Hierarchical EVC (H-EVC) models. The original formulation primarily addressed discrete, single-step decisions occurring within structured, isolated experimental trials. However, ecological human behavior is inherently hierarchical, spanning multiple embedded timescales: an individual may mobilize control to hit an isolated keystroke, in service of writing a scientific paper, in service of completing a doctoral degree, in service of advancing a lifelong career.

Hierarchical EVC frameworks, advanced by researchers such as Amitai Shenhav, Sebastian Musslick, and their collaborators, model control allocation across deep temporal and representational hierarchies. In an H-EVC architecture, high-level control nodes (situated within frontomedial and rostrolateral prefrontal structures) specify abstract, long-term policy goals and distribute intermediate control budgets to lower-level nodes. These lower-level nodes (situated within caudal dACC and premotor regions) optimize immediate, fine-grained control signals targeting concrete sensorimotor streams.

This hierarchical nesting introduces the concept of meta-control: recursive computational loops where supervisory control policies are deployed to regulate, monitor, and optimize other control policies. Meta-control algorithms dynamically determine when the brain should pause to reconsider its global strategy versus when it should execute established sub-routines automatically. This expansion bridges the gap between basic trial-by-trial interference tasks and the complex, multi-scale problem-solving that defines extended human agency.

11.2 The dACC Specialization Controversy

Despite the unifying success of the EVC framework, its anatomical assignment of this optimization engine specifically to the dorsal anterior cingulate cortex has generated substantial controversy and critical pushback from empirical neurophysiologists. Chief among the critics are researchers who argue that characterizing the dACC as a high-level cognitive “evaluative optimizer” over-intellectualizes a structure whose evolutionary origins are fundamentally premotor and autonomic.

Motor physiologists, such as Paul Cisek, point out that the cingulate cortex is phylogenetically ancient, possessing dense, direct connections to motor effectors and autonomic centers long before the evolutionary expansion of the granular prefrontal cortex in anthropoid primates. They argue that cingulate activations during cognitive control tasks do not reflect an abstract economic calculation; rather, they reflect concrete, low-level competition between rival motor programs, behavioral switching, and visceral cardiovascular preparation for immediate physical action. In this view, framing the dACC as an economic executive homunculus risks obscuring its grounded evolutionary role as a somatic-motor selection engine.

Simultaneously, the foraging controversy spearheaded by Nils Kolling, Matthew Rushworth, and their colleagues sparked a major theoretical debate over functional specialization. Kolling and colleagues published empirical fMRI studies arguing that the dACC’s activation during difficult choices does not reflect the expected value of control, but rather the relative value of *foraging*—specifically, the value of the unchosen options versus the default choice. Shenhav and Cohen contested this interpretation through extensive re-analyses and formal computational modeling, demonstrating that the foraging-value signals identified by Kolling were confounded by choice difficulty, decision conflict, and the rising demand for mental effort. This ongoing debate has prompted both sides to design increasingly refined paradigms, ultimately confirming that while the dACC undoubtedly interfaces directly with motor and autonomic outputs, its activity profile reliably matches the multiplexed signatures of an integrated cost-benefit optimizer.

11.3 Bounded Rationality and Heuristic EVC Approximations

Another major critique leveled against early formulations of EVC concerns the issue of computational intractability. Critics pointed out that the fully specified EVC objective function:
$$arg\max_{signal} \left[ \sum_{i} P(Outcome_i mid S, signal) \cdot U(Outcome_i) – \text{Cost}(signal) \right]$$
demands calculations that are themselves computationally demanding. If the brain must evaluate complex, multi-dimensional probability distributions and calculate gradients across thousands of potential control parameters on a millisecond timescale, the meta-level cost of computing the optimal control allocation could easily exceed the benefits of the control itself. The brain would fall victim to an infinite regress of meta-control: allocating control to optimize the control that optimizes control.

To resolve this tension, contemporary EVC models are explicitly grounded within the principles of bounded rationality and resource-rational analysis, frameworks championed by Herbert Simon and extended to computational neuroscience by researchers like Falk Lieder and Thomas Griffiths. Under a resource-rational EVC model, the supervisory controller does not attempt to calculate the analytically perfect, globally optimal control allocation. Instead, it seeks a satisficing solution: an approximation that achieves good-enough performance while minimizing the computational cost of the optimization process itself.

These satisficing policies are realized through algorithmic simplifications, such as sampling-based approximations (e.g., Monte Carlo policy evaluations with very few samples), coarse discretization of control intensities into binary or ternary levels, and the heavy use of amortized meta-learning. Under severe time pressure, the EVC system relies on simple heuristics—such as “if conflict is detected, bump control intensity by a fixed increment”—and only engages deeper, iterative computational optimizations when environmental stakes are exceptionally high and temporal deadlines allow. This integration of bounded rationality protects the EVC model from computational paralysis, grounding it firmly within the finite processing capabilities of biological nervous systems.

12. Synthesis, Future Horizons, and Unresolved Questions in EVC Research

12.1 Laminar fMRI and Optogenetic Testing of Microcircuits

The continuous evolution of neuroscience methodology provides unprecedented opportunities to test the mechanistic predictions of EVC theory at cellular and laminar resolution. Historically, human cognitive neuroimaging relied on standard 3-Tesla fMRI, which averages blood oxygenation signals across several cubic millimeters of brain tissue—blurring together the activity of hundreds of thousands of heterogeneous neurons across all cortical layers. This spatial blurring prevented researchers from conclusively distinguishing between feedforward input signals, local computational transformations, and descending efferent control commands within the dACC.

The advent of ultra-high-field (7-Tesla and 9.4-Tesla) laminar fMRI has fundamentally altered this landscape. Laminar neuroimaging allows researchers to dissociate hemodynamic responses occurring within the superficial layers (layers I–III), the middle input layers (layer IV), and the deep output layers (layers V–VI) of the human cingulate cortex. EVC theory makes precise, falsifiable predictions regarding this laminar organization:

• Superficial and intermediate layers should predominantly receive and process multimodal feedforward inputs—such as value signals from the vmPFC, interoceptive signals from the insula, and conflict signals from sensory cortices.
• Deep pyramidal layers (layers V and VI), which project directly to the dlPFC, striatum, locus coeruleus, and cingulate motor areas, should specifically reflect the resolved output of the EVC calculus: the specified identity and intensity of the chosen control vector.

Concurrently, in animal models, the deployment of cell-type-specific optogenetics, chemogenetics (DREADDs), and two-photon calcium imaging enables causal, mechanistic manipulations of these circuits. By selectively exciting or silencing specific genetically tagged neuronal subpopulations within the rodent or non-human primate cingulate cortex—such as parvalbumin-positive inhibitory interneurons or specific corticostriatal projection neurons—researchers can directly alter the internal cost-benefit computation. Causal optogenetic tests have demonstrated that artificial stimulation of specific dACC assemblies can induce animals to persist through high physical and cognitive effort barriers to secure rewards, offering proof of the causal role of these microcircuits in setting the subjective value of mental effort.

12.2 Applications to Artificial Intelligence and Autonomous Agents

The conceptual framework of EVC has also expanded into the domains of artificial intelligence, computational robotics, and neuromorphic engineering. Contemporary deep reinforcement learning (DRL) systems, while capable of superhuman performance in closed domains like chess, Go, and video games, remain notoriously brittle, computationally inefficient, and prone to catastrophic resource saturation. Unlike biological organisms, state-of-the-art artificial neural networks rarely possess an internal, explicit metric for the cost of their own computations. They expend identical compute budgets processing trivial visual scenes as they do navigating complex, high-stakes environments, leading to extreme energy consumption and vulnerabilities in dynamic real-world environments.

Integrating EVC principles into AI architectures offers a principled pathway toward resource-rational autonomous agents. By embedding an explicit “computational cost penalty” into the agent’s meta-level loss function, artificial agents can learn to dynamically modulate their own internal inference budgets. An EVC-guided artificial agent scales its compute—allocating more Monte Carlo rollouts, deeper recurrent passes, or larger transformer attention spans—only when the incoming environmental state presents high ambiguity, danger, or reward potential.

Furthermore, this dynamic is critical in the development of energy-efficient neuromorphic hardware. Biological brains operate within a strict thermal and metabolic envelope, consuming roughly 20 watts of power. They achieve this extraordinary efficiency by relying primarily on cheap, sparse, automatic processing, strategically reserving high-intensity, synchronized neural firing for situations where the expected value of intervention is demonstrably high. By designing neuromorphic chips that implement EVC-style control gating, computer scientists can engineer autonomous machines, drones, and edge-computing devices capable of surviving and optimizing their processing in complex environments under stringent battery and compute constraints.

12.3 Unifying Value, Effort, and Executive Cognition

Looking back across more than a decade of research since its introduction, the Expected Value of Control theory by Amitai Shenhav, Matthew M. Botvinick, and Jonathan D. Cohen stands as a major achievement in modern cognitive neuroscience. Prior to this landmark formulation, the field was divided: cognitive control was conceptualized as a cold, mechanistic set of computational filters, while neuroeconomics and motivational psychology operated as warm, value-driven disciplines that rarely probed the computational bottlenecks of executive microcircuits. The anterior cingulate cortex remained an empirical battleground, torn between competing paradigms of conflict, error, pain, and emotion.

The EVC framework dismantled these divisions. By recasting cognitive control as an optimal, resource-constrained economic decision, it integrated executive function, bounded rationality, sequential evidence accumulation, and reinforcement learning into a single mathematical objective function. It provided an elegant, computationally parsimonious account of anterior cingulate function, demonstrating how physical pain, emotional distress, motor error, and computational conflict all act as integrated parameters feeding an overarching valuation: determining whether, where, and how intensely to invest the brain’s limited mental labor.

As cognitive neuroscience advances into the coming decades, persistent theoretical questions remain. Researchers must continue to delineate the precise laminar and cellular mechanisms mediating the EVC computation, resolve the dynamic arbitration between model-based and heuristic control strategies, unravel how developmental trajectories and psychiatric pathologies warp effort cost functions, and translate these biological insights into resource-rational artificial intelligence. Yet, whatever refinements and paradigm shifts lie ahead, the conceptual core established by Shenhav, Botvinick, and Cohen remains foundational: cognitive control is not a detached, homuncular supervisor operating outside the laws of economics, but a value-driven, cost-sensitive optimization process that balances the exertion of mental labor against the rich rewards of adaptive thought and action.

References

• Aston-Jones, G., & Cohen, J. D. (2005). An integrative theory of locus coeruleus-norepinephrine function: Adaptive gain and optimal performance. Annual Review of Neuroscience, 28, 403–450. https://doi.org/10.1146/annurev.neuro.28.061604.135709
• Baumeister, R. F., Bratslavsky, E., Muraven, M., & Tice, D. M. (1998). Ego depletion: Is the active self a limited resource? Journal of Personality and Social Psychology, 74(5), 1252–1265. https://doi.org/10.1037/0022-3514.74.5.1252
• Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S., & Cohen, J. D. (2001). Evaluating the demand for control: Executive function and the anterior cingulate cortex. Psychological Review, 108(3), 624–652. https://doi.org/10.1037/0033-295X.108.3.624
• Cason, C. A., & Egner, T. (2017). The learned value of cognitive control: How targeted reinforcement shapes executive allocation. Cognitive, Affective, & Behavioral Neuroscience, 17(5), 980–994. https://doi.org/10.3758/s13415-017-0527-2
• Cavanagh, J. F., & Frank, M. J. (2014). Frontal theta as a mechanism for cognitive control. Trends in Cognitive Sciences, 18(8), 414–421. https://doi.org/10.1016/j.tics.2014.04.012
• Cisek, P. (2007). Cortical mechanisms of action selection: The affordance competition hypothesis. Philosophical Transactions of the Royal Society B: Biological Sciences, 362(1485), 1585–1599. https://doi.org/10.1098/rstb.2007.2054
• Daw, N. D., Niv, Y., & Dayan, P. (2005). Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nature Neuroscience, 8(12), 1704–1711. https://doi.org/10.1038/nn1560
• Duncan, J. (2010). The multiple-demand (MD) system: A key framework for understanding cognitive capacity and executive control. Neuropsychologia, 48(13), 3467–3478. https://doi.org/10.1016/j.neuropsychologia.2010.08.019
• Gehring, W. J., Goss, B., Coles, M. G., Meyer, D. E., & Donchin, E. (1993). A neural system for error detection and compensation. Psychological Science, 4(6), 385–390. https://doi.org/10.1111/j.1467-9280.1993.tb00586.x
• Gratton, G., Coles, M. G., & Donchin, E. (1992). Optimizing cognitive processing: A model of dynamic regulation of executive function and attentional selection. Journal of Experimental Psychology: General, 121(4), 480–506. https://doi.org/10.1037/0096-3445.121.4.480
• Inzlicht, M., Schmeichel, B. J., & Macrae, C. N. (2014). Why self-control seems (but may not be) limited. Trends in Cognitive Sciences, 18(3), 127–133. https://doi.org/10.1016/j.tics.2013.12.009
• Kogler, L., Müller, V. I., Chang, A., Eickhoff, S. B., Fox, P. T., & Gur, R. C. (2015). Cognitive control in the anterior cingulate cortex: Common and distinct networks for conflict and task difficulty. NeuroImage, 120, 294–306. https://doi.org/10.1016/j.neuroimage.2015.07.018
• Kolling, N., Behrens, T. E., Mars, R. B., & Rushworth, M. F. (2012). Neural mechanisms of foraging in the human medial prefrontal cortex. Science, 336(6077), 95–98. https://doi.org/10.1126/science.1216930
• Kool, W., McGuire, J. T., Rosen, Z. B., & Botvinick, M. M. (2010). Decision making and the avoidance of cognitive demand. Journal of Experimental Psychology: General, 139(4), 665–682. https://doi.org/10.1037/a0020198
• Lieder, F., & Griffiths, T. L. (2020). Resource-rational analysis: Understanding human cognition as the optimal use of bounded computing resources. Behavioral and Brain Sciences, 43, e1. https://doi.org/10.1017/S0140525X1900151X
• Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167–202. https://doi.org/10.1146/annurev.neuro.24.1.167
• Musslick, S., & Cohen, J. D. (2021). Rationalizing constraints on the capacity for cognitive control. Nature Reviews Neuroscience, 22(12), 757–773. https://doi.org/10.1038/s41583-021-00516-z
• Norman, D. A., & Shallice, T. (1986). Attention to action: Willed and automatic control of behavior. In R. J. Davidson, G. E. Schwartz, & D. Shapiro (Eds.), Consciousness and Self-Regulation: Advances in Research and Theory (pp. 1–18). Springer. https://doi.org/10.1007/978-1-4757-0629-1_1
• Ratcliff, R., & McKoon, G. (2008). The diffusion decision model: Theory and data for two-choice decision tasks. Neural Computation, 20(4), 873–922. https://doi.org/10.1162/neco.2008.12-06-420
• Shackman, A. J., Salomons, T. V., Slagter, H. A., Fox, A. S., Winter, J. J., & Davidson, R. J. (2011). The integration of negative affect, pain and cognitive control in the cingulate cortex. Nature Reviews Neuroscience, 12(3), 154–167. https://doi.org/10.1038/nrn2994
• Shenhav, A., Botvinick, M. M., & Cohen, J. D. (2013). The expected value of control: An integrative theory of anterior cingulate cortex function. Nature Reviews Neuroscience, 14(3), 217–226. https://doi.org/10.1038/nrn3564
• Shenhav, A., Cohen, J. D., & Botvinick, M. M. (2016). Dorsal anterior cingulate cortex and the value of control. Nature Neuroscience, 19(10), 1286–1291. https://doi.org/10.1038/nn.4384
• Shenhav, A., Musslick, S., Lieder, F., Kool, W., Griffiths, T. L., Cohen, J. D., & Botvinick, M. M. (2017). Toward a rational and mechanistic account of mental effort. Annual Review of Neuroscience, 40, 99–124. https://doi.org/10.1146/annurev-neuro-072116-031526
• Simon, H. A. (1955). A behavioral model of rational choice. The Quarterly Journal of Economics, 69(1), 99–118. https://doi.org/10.2307/1884852
• Treadway, M. T., Buckholtz, J. W., Cowan, R. L., Woodward, N. D., Li, R., Ansari, M. S., Baldwin, R. M., Schwartzman, A. N., Kessler, R. M., & Zald, D. H. (2012). Dopaminergic mechanisms of individual differences in human effort-based decision-making. Journal of Neuroscience, 32(18), 6170–6176. https://doi.org/10.1523/JNEUROSCI.6459-11.2012
• Westbrook, A., & Braver, T. S. (2015). Cognitive effort: A neuroeconomic approach. Cognitive, Affective, & Behavioral Neuroscience, 15(2), 395–415. https://doi.org/10.3758/s13415-015-0334-y
• Westbrook, A., Kester, D., & Braver, T. S. (2013). What is the subjective cost of cognitive effort? Load, preference, and the discounted value of reward. PLOS ONE, 8(7), e68210. https://doi.org/10.1371/journal.pone.0068210