Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Goal-Directed Action and Habit Formation

Abstract. Goal-directed action and habit formation are complementary forms of learned behavioral control. This fundamental research monograph examines how anticipated outcomes, causal action–outcome contingencies, contextual cues, and repetition influence action selection and its flexibility. A critical narrative synthesis of historical instrumental-conditioning experiments, rodent causal interventions, human neuroimaging and behavioral tasks, computational reinforcement-learning models, and longitudinal everyday-habit research distinguishes outcome devaluation from contingency degradation, learned automaticity from response frequency, and laboratory habits from compulsive clinical behavior. The account evaluates corticostrial circuitry, model-based and model-free algorithms, stress-related modulation, behavioral interventions, and measurement validity. Contradictory human overtraining results, 2025–2026 preregistered replication and convergent-validity evidence, task reliability, model identifiability, and the boundary between mechanistic and functional explanation are examined explicitly. The synthesis identifies robust action–outcome sensitivity and causal neural evidence, conditional evidence for human habit expression, and priorities for reproducible multimethod research. Artificial control systems are discussed as computational comparators within the Artificial Era. This work is a critical scholarly narrative synthesis rather than a registered systematic review, original experiment, or meta-analysis.

Keywords: goal-directed action; habit formation; instrumental conditioning; outcome devaluation; contingency degradation; action–outcome association; stimulus–response habit; model-based reinforcement learning; model-free reinforcement learning; corticostriatal circuits; behavioral automaticity; habit measurement; compulsion; behavior change

The Problem of Action Control

Goal-directed action is the selection and performance of a response under the control of anticipated consequences. A goal-directed agent can adjust what it does when a particular consequence loses value or when the response ceases to produce that consequence. Habitual control describes learned tendencies for contextual stimuli to recruit responses with relative independence from the current evaluation of their results. These definitions identify testable control relations rather than personality types or everyday descriptions of determination and routine. The same observable action may be generated by either process, and an individual can use both in different proportions across trials. The foundational experimental task is therefore to identify the variables governing response performance, not to classify an action by its appearance or speed. Balleine & Dickinson (1998) Balleine & O’Doherty (2010)

Instrumental behavior connects an organism's actions to environmental changes. A person can choose a route because it is expected to reach a valued destination, then traverse familiar turns almost automatically after repetition. The route example reveals an identification problem: successful performance does not by itself indicate whether consequences are currently represented. Learning may establish a response–outcome association, a stimulus–response tendency, contextual predictions, and motor skills together. A carefully designed experiment changes the value or causal effectiveness of the outcome while keeping essential cues available. Performance following that intervention provides evidence about the controlling process, although a negative result requires further checks that the manipulation was actually learned and remembered. Adams & Dickinson (1981) de Wit & Dickinson (2009)

The research program joins comparative animal learning, cognitive theories of action, computational decision making, behavioral neuroscience, health psychology, and research on everyday automaticity. Its central questions concern acquisition, representation, expression, arbitration, and change. Acquisition asks what a learner encodes; expression asks which information determines the response on a particular occasion. Arbitration asks how several policies compete or cooperate. Change asks when outcome updating, contextual alteration, or new practice shifts observable behavior. These questions are related but experimentally separable. A complete synthesis examines classical response-contingency experiments, causal manipulations of brain systems, human behavioral paradigms, and prospective real-world data. Distinguishing their respective inferential reach prevents clinical or social generalizations from outrunning laboratory results. Wood & Rünger (2016) de Wit et al. (2018)

A Historical Reconstruction of Actions and Habits

Classical instrumental-conditioning research established that consequences modify the probability of subsequent responding, but the resulting performance left open what was learned about the action itself. A reinforcement history might strengthen a direct link from situation to response, or it might teach a representation of the particular effect the response causes. Outcome devaluation provided a decisive experimental strategy: train a specific action for a particular reward, change the reward's value independently, and measure the trained response when the original reward is withheld. The landmark experiments of Adams & Dickinson (1981) showed that animals could reduce performance after the relevant reinforcer was devalued. This effect furnished behavioral evidence for knowledge of an action's specific consequence beyond a simple accumulated reinforcement count.

The 1980s and 1990s elaborated a distinction between behavioral autonomy after prolonged training and control based on action–outcome knowledge. Subsequent work specified two components of goal direction: knowledge that an action produces an identifiable outcome and learning the current incentive value of that outcome. Balleine & Dickinson (1998) linked contingency and incentive learning to experimental preparations and cortical substrates. The distinction corrected an ambiguity of devaluation results: reducing a response after a food becomes aversive presupposes both an association with the food and the ability to deploy its changed value. Likewise, response persistence may reflect incomplete devaluation, performance difficulty, or contextual interference as well as habitual control. Historical development made task manipulation and interpretation inseparable.

Research on human action added ideomotor traditions in which representing an intended effect can activate responses previously associated with producing it. de Wit & Dickinson (2009) argued that associations can support several temporal orders of decision: desired outcome may prime an action, while considering an action may call its consequence to mind for evaluation. Neuroimaging and computational reinforcement learning then extended the theory beyond the simple opposition of two independent mechanisms. Daw et al. (2005) formalized how reliability could influence allocation of control, and Wood & Rünger (2016) situated laboratory habit theory alongside repeated behavior in stable everyday contexts. The historical sequence is a progression of operational distinctions and converging methods, with continuing disagreement over how strongly the classical dissociation generalizes to humans.

Operational Definitions and Distinct Constructs

An action–outcome association encodes a conditional relation between performance and the identity of its consequence. Incentive value concerns the consequence's current motivational significance, which changes with satiation, deprivation, aversion, or changing objectives. An antecedent stimulus is a discriminative or contextual feature present before a response; stimulus–response learning can increase the probability of a response in that context. Habit strength is a latent disposition to recruit an acquired response under relevant cues. Behavioral automaticity denotes experiential characteristics such as low deliberative demand, initiation without explicit intention, and relative efficiency. These constructs overlap in ordinary use but require distinct measurement. Equating repetition frequency with automaticity eliminates the very hypothesis that practice can change response control. Verplanken & Orbell (2003) Gardner et al. (2012)

Goal-directedness is an empirical property of an action in a specified testing situation. The strongest classical evidence combines outcome-value sensitivity with action–outcome contingency sensitivity. A participant may continue a behavior after a reward loses value because the value change did not transfer to the test context, because the response was initiated before updating, or because another valued consequence remains. Conversely, stopping a response after devaluation can reflect an explicit instruction, response inhibition, or a decision rule rather than a stable goal-based policy acquired during training. The measurement target must therefore specify what was trained, the devaluation operation, whether outcome identity was recalled, and what alternative control pathways remain available. Balleine & O’Doherty (2010) de Wit et al. (2018)

Three additional distinctions organize this field. Acquisition and expression differ because a trained habit can remain latent when deliberate planning wins response selection; behavior may reveal a habit only under restricted preparation time. Skill differs from habit because accurate automated execution says little about sensitivity to valued outcomes. Compulsion differs from habit because clinical repetition also implicates distress, impairment, conflict, and motivational factors not captured by a simple devaluation contrast. Hardwick et al. (2019) demonstrates the importance of response-preparation timing; Everitt & Robbins (2016) analyzes relationships among goal-directed action, habitual control, and compulsive drug seeking. Treating these as separate explananda supports comparison without assuming a single continuum or mechanism.

Action–Outcome Identity and Causal Contingency

Goal-directed control requires a representation of what an action tends to produce and how that production depends on acting. Outcome identity matters because two actions may earn rewards that are equally pleasant yet differ in sensory, social, or instrumental properties. A learner who distinguishes the consequences can adjust the relevant action selectively when one reward changes. Causal contingency also matters: a response loses its instrumental usefulness when the outcome arrives just as frequently without the response. Contiguity alone cannot solve this problem, since outcomes may occur close to actions even when their occurrence is statistically independent. Balleine & Dickinson (1998) identifies action–outcome contingency and incentive learning as joint determinants.

The contingency of an action can be expressed as the difference between the probability of the outcome given performance and the probability of that outcome without performance, conditional on the experimental setting. Interventions that deliver the formerly earned outcome independently of response performance reduce this difference while preserving exposure to the outcome. A decrease in the trained response under such degradation indicates sensitivity to its instrumental effectiveness. A persistent response suggests weaker control by current contingency, provided learning about the degradation and the availability of relevant alternatives have been verified. Studies using dorsal-striatal interventions demonstrate that contingency degradation and devaluation can be affected together, strengthening the inference that action-based representations are involved. Yin et al. (2005)

An action's consequences are rarely unitary in natural settings. A habitual morning walk may produce exercise, social contact, relaxation, and progress toward a larger objective; removing one reward may leave other consequences valued. In addition, an action can alter future opportunities rather than immediately deliver an outcome. Research tasks often simplify this structure to one or two explicitly identifiable rewards, increasing internal validity while reducing ecological breadth. A comprehensive theory must therefore model both the action–outcome mapping actually learned in the experiment and the potentially richer consequence structure of everyday action. Contingency degradation is especially informative when reward frequency is controlled and discrimination checks establish that participants noticed the manipulated causal relation. de Wit & Dickinson (2009) Wood & Rünger (2016)

Incentive Learning and Outcome Revaluation

Reward devaluation changes the motivational value of an outcome after its instrumental association has been learned. Common procedures include sensory-specific satiation and conditioned taste aversion in animal preparations and selective satiety or instruction-supported revaluation in human tasks. If the subject suppresses the action that formerly yielded the devalued outcome, the result supports the interpretation that current outcome value contributes to control. A test conducted in extinction prevents immediate new reinforcement from explaining the difference: choice must rely on previously acquired action–outcome knowledge together with updated value. Adams & Dickinson (1981) pioneered this logic, and Valentin et al. (2007) used food-specific satiation during human neuroimaging to examine value-sensitive action selection.

Devaluation is a sequence of interventions, each of which requires its own manipulation check. First, the participant must learn which action yields which consequence. Next, the value intervention must measurably change the consequence's desirability. Finally, the test must permit expression of the learned association without overwhelming response constraints. When devaluation is achieved by satiety, general fatigue or reduced motivation may suppress every response rather than the specific devalued one. When it is achieved by instruction, participants may follow an explicit stopping rule. Matched control outcomes, outcome-specific preferences, extinction timing, and memory probes are therefore essential to identifying the causal pathway from changed value to action selection. Balleine & O’Doherty (2010) de Wit et al. (2018)

An important theoretical implication is that goal-directed action is sensitive to a valuation updated outside the original acquisition episode. This property makes the assay more revealing than whether performance improves during training. However, a revaluation result does not identify a unique representational format: an associative chain, explicit model, prospective simulation, or learned conditional rule may all mediate change. Neuroimaging can identify correlated value signals, but inference about necessity requires causal manipulation or convergent lesion evidence. The empirical claim is that present outcome value exerts control under well-specified conditions; a further model is needed to explain the representations and computations by which that influence occurs. de Wit & Dickinson (2009) Valentin et al. (2007)

Contingency Degradation as an Experimental Intervention

Contingency degradation asks whether a learned action is still produced when its distinctive causal contribution to obtaining a reward has weakened. In an instrumental preparation, the experimenter can schedule additional outcome deliveries in the absence of the previously rewarded response, thereby reducing the advantage of acting. A response that declines selectively under this intervention is sensitive to action effectiveness. The measure is especially valuable when the outcome remains attractive, because it targets causal knowledge rather than incentive value. Balleine & Dickinson (1998) established the necessity of separating these dimensions, and Yin et al. (2005) showed how interventions in posterior dorsomedial striatum can alter both contingency and value sensitivity in rats.

The comparison must account for how often the animal or participant experiences response-contingent and noncontingent rewards, the response opportunity, context, baseline performance, and possible temporal contiguity. An increase in freely delivered outcomes can cause satiety or alter arousal; a corresponding reduction in response would then have an interpretation unrelated to causal degradation. Researchers can include controls for outcome rate, probe alternative actions, and compare intact versus manipulated associative relationships. In tasks with several actions, selectively degrading one mapping is particularly informative because a general decline in reward motivation would tend to affect both. The central inferential ambition is to isolate learning that an action makes a difference to a particular outcome. Balleine & O’Doherty (2010)

Contingency manipulations are difficult to transport directly into human everyday settings. Natural responses often influence outcomes indirectly, with long delays and multiple contributing agents. A person may keep performing a routine despite low current effectiveness because it signals identity, coordinates with others, or maintains a downstream opportunity. Thus, laboratory contingency insensitivity demonstrates a specific restriction of flexible action control under specified conditions, while general claims of irrationality require a wider model of rewards and constraints. The most defensible translational design combines a precise laboratory intervention with field measurements of perceived effectiveness, actual outcome availability, and context recurrence. de Wit & Dickinson (2009) Wood & Rünger (2016)

Training Duration, Reinforcement Schedules, and Practice

Repeated performance in a stable stimulus environment is a major candidate mechanism of habit acquisition. Each episode may strengthen a context-linked response tendency while also providing new information about the action's consequences. Classical studies demonstrated that sufficiently extended instrumental training can produce behavior less sensitive to outcome devaluation, but the rate of transition depends on reinforcement schedules, choice structure, and outcome representation. In interval schedules, reward becomes available with time rather than strictly proportional to each response; this can weaken the informational connection between action count and outcome delivery. Yin et al. (2004) used such training to distinguish intact overtrained habits from effects of dorsolateral striatal lesions.

Repetition must be analyzed at several levels. The number of trials represents exposure, whereas reinforcement schedules determine which transitions are experienced and the informativeness of feedback. Stable discriminative stimuli provide retrieval cues, while changes in response cost and delay influence whether a cached policy remains useful. Overtraining may also improve motor proficiency, response speed, and confidence, so a shorter reaction time after practice provides insufficient evidence of outcome-insensitive control. A well-designed training experiment compares multiple doses, equates immediate performance when possible, and administers the same sensitive devaluation or contingency test across doses. Tricomi et al. (2009) supplied a human free-operant example; de Wit et al. (2018) documented the practical difficulty of producing comparable transitions reliably.

Practice may foster both a flexible policy and a fast alternative policy at the same time. An increase in the habitual candidate's strength does not require immediate visible domination of behavior. Hardwick et al. (2019) found that trained responses could be revealed under short preparation deadlines even after participants could deliberately follow a remapped response rule. This result suggests that the time available for selection mediates the relation between acquired tendencies and expressed actions. Accordingly, overtraining experiments should measure response initiation as well as final choice, vary time pressure explicitly, and avoid treating the absence of errors under leisurely test conditions as evidence that no habitual tendency was learned.

Extinction Tests and Performance Interpretation

An extinction test withholds previously available rewards while measuring whether a trained response is selected. When it immediately follows an outcome-value manipulation, extinction limits the opportunity for fresh reinforcement during the critical probe and helps identify the use of previously learned information. In a well-controlled devaluation assay, a decrease for the devalued action relative to a still-valued action constitutes evidence of outcome-sensitive choice. The experimenter must nevertheless account for the fact that extinction itself is new learning: after repeated unrewarded trials, performance can change for reasons independent of devaluation. The order, duration, and separation of probes should therefore be designed to minimize this confounding process. Adams & Dickinson (1981) Balleine & Dickinson (1998)

Response counts under extinction summarize performance and are subject to floor and ceiling effects. A participant who rarely responds under any condition may appear equally insensitive to the two outcome values, even when the underlying action policy is flexible. Conversely, an individual who makes very few errors may appear entirely controlled by goals despite possessing a strong latent habitual tendency. Task difficulty, monetary or food incentives, response time, working-memory demand, and explicit rule comprehension all affect the measurement range. Mixed-effects analyses can estimate within-participant contrasts across outcome value and training dose, while preregistration can specify the critical interaction rather than selecting an attractive post hoc index. de Wit et al. (2018) Hardwick et al. (2019)

Interpreting null devaluation differences is especially demanding. Statistical nonsignificance is not positive evidence of habitual control; the assay must first demonstrate that devaluation succeeded and that the test has adequate precision to reveal a relevant difference. Evidence for habit is stronger when the design establishes prior action–outcome learning, verifies the value manipulation, includes a positive goal-directed control, and documents response persistence under an appropriate outcome-independent probe. Equivalence or interval-based tests may help quantify how much sensitivity can reasonably be excluded, but only if meaningful bounds and adequate power are specified. These safeguards make a negative result an interpretable measurement of action control rather than an artifact of failed manipulation. de Wit et al. (2018) Luque et al. (2020)

Outcome-Specific Transfer and Cue-Driven Performance

Learned cues can influence actions through several pathways. A discriminative stimulus may signal that a response will be reinforced; a Pavlovian cue may acquire motivational influence because it predicts a valued outcome; and repeated co-occurrence may establish a direct stimulus–response tendency. Outcome-specific Pavlovian-to-instrumental transfer investigates whether a stimulus predicting a particular outcome selectively promotes an action trained to obtain the same outcome. The resulting effect demonstrates interaction among cue, outcome identity, and action selection. It does not, by itself, establish that the action is habitual: a cue could recruit an outcome representation that participates in flexible selection. Balleine & O’Doherty (2010) de Wit & Dickinson (2009)

Distinguishing direct response triggering from mediated outcome expectancy requires a factorial approach to cues and rewards. Researchers can devalue one outcome, test transfer separately from instrumental response selection, and determine whether cues continue to bias responding under altered contingencies. If a stimulus can influence choice despite a reduced instrumental value, the observation must still be compared with alternative accounts involving incentive salience, retrieval, attention, and imperfect representation of the changed value. Causal analyses should specify whether the proposed effect is a change in stimulus-driven action activation, an increase in general motivational vigor, or a shift in the relative valuation of available outcomes. These may coexist, but their experimental signatures differ. Balleine & Dickinson (1998) Everitt & Robbins (2016)

Everyday environments are dense with cues that recur at particular times, locations, and social occasions. This stability can make previously selected actions easier to initiate, enabling efficient routine performance. It can also produce context-dependent recurrence when a person intends to change behavior but encounters the original setting. Field research measures how frequently contexts and behaviors co-occur, while laboratory tasks identify the mechanisms by manipulating cues and expected consequences. Their combination supports a grounded account of habit expression: contextual recurrence increases response availability, and the action ultimately observed depends on value, competing goals, opportunity, and time for correction. Wood & Rünger (2016) Hardwick et al. (2019)

Comparative Species and Translational Inference

Rodent experiments contribute strong causal evidence because action outcomes, training schedules, anatomical interventions, and subsequent test conditions can be controlled precisely. The classic dissociations of dorsomedial and dorsolateral striatal involvement become compelling when lesions or reversible inactivation alter devaluation sensitivity while preserving important aspects of basic response performance. Yin et al. (2004) and Yin et al. (2005) helped distinguish neural support for habitual and goal-directed control. However, these effects are conditional on the trained species, task, precise subregion, stage of learning, and method of intervention. A striatal lesion changes an animal's neural capacities; its behavioral effect is evidence about the necessity of a circuit in that preparation rather than a complete mapping of psychological processes to one anatomical site.

Human studies provide direct access to task instructions, explicit beliefs, verbal explanations, subjective reward valuations, and broad everyday behavior. They also introduce strategic compliance, rapid task abstraction, and the possibility that an instruction-based response bypasses the learning process meant to be measured. Neuroimaging commonly records correlational blood-oxygen-level-dependent signals with regional and temporal limits, whereas behavioral assays can identify an effect without localizing a neural implementation. Valentin et al. (2007) found value-sensitive orbitofrontal activity in a human food-choice task, and Tricomi et al. (2009) related practice to posterior putamen activity and outcome sensitivity. Their conclusions have distinct inferential scopes.

A rigorous cross-species synthesis aligns the manipulation and outcome variable before comparing neural systems. The same phrase, habit learning, can denote automatic performance, insensitivity to devaluation, high practice dose, or a computationally cached policy in different studies. Only some of these provide the classical instrumental definition. Functional similarities across rodents and humans can motivate hypotheses about homologous corticostriatal circuits without proving one-to-one correspondence of psychological mechanisms. Balleine & O’Doherty (2010) reviews evidence for comparative homologies while recognizing integrative circuitry. Future translational studies should combine harmonized devaluation measures, contingency tests, response-time constraints, and training histories, thereby testing common functional predictions rather than transferring labels alone.

Dorsomedial Striatum and Goal-Directed Control

The dorsomedial striatum participates in circuitry supporting selection based on learned action consequences. In rats, Yin et al. (2005) compared lesions of posterior and anterior dorsomedial regions, posttraining interventions, and reversible inactivation while assessing both outcome devaluation and contingency degradation. The posterior-region manipulations impaired sensitivity to these interventions under conditions in which basic response and outcome discriminations could remain available. This combination is important because a generalized movement impairment would not specifically explain failure to adapt action selection to new value or contingency information. The results support a causal contribution to acquisition and expression of goal-directed instrumental control, although the exact computations carried by the subregion remain a further question.

Prefrontal and corticostriatal interactions provide a broader network context for this dissociation. The prelimbic cortex has been implicated in acquiring action–outcome relationships, while orbitofrontal and related cortical territories contribute to outcome evaluation and flexible selection in task-dependent ways. Balleine & O’Doherty (2010) compares candidate functional homologies across species and emphasizes the integration of motivational information with selection. Anatomical nomenclature alone does not establish a functional homology: species differ in cortical organization, and a lesion or signal in one location can affect connected regions. A strong account therefore uses behaviorally defined devaluation and degradation effects as the reference points and interprets the circuitry relative to those tested functions.

Human neuroimaging extends this work by examining valuation changes while people choose among learned actions. In Valentin et al. (2007), food outcomes were separately associated with choices before one was devalued through selective satiety, and orbitofrontal activity changed during subsequent selections. The data support a role for orbitofrontal valuation processes in flexible action. BOLD activity, however, cannot by itself identify an indispensable neural substrate, exclude signals reflecting attention or visual properties, or establish the exact association encoded by neurons. The convergent interpretation combines rat intervention studies with human value-sensitive signals, recognizing that the two methods answer different causal and representational questions.

Dorsolateral Striatum and Habit Expression

Lesions of dorsolateral striatum can preserve outcome expectancy while disrupting the acquisition or expression of outcome-insensitive habits after extensive instrumental training. Yin et al. (2004) reported that lesioned rats remained responsive to devaluation under schedules that produced relatively habitual performance in controls. This causal contrast supports a role for dorsolateral striatal circuits in learned stimulus-linked responding. It also illustrates the distinction between learning what a reward is and using that information to govern an action: an animal may discriminate reward identities yet act through a policy whose current expression is less responsive to incentive revaluation. The circuit's contribution should be interpreted within a broader network of sensorimotor, thalamic, and cortical pathways.

In humans, Tricomi et al. (2009) investigated a free-operant learning task with different amounts of practice and later devaluation. Greater training coincided with reduced behavioral sensitivity to altered outcome value, while cue-related fMRI responses increased in a posterior putamen–globus pallidus region. The findings constitute an influential demonstration linking an instrumental habit criterion with a candidate neural correlate in humans. The design nevertheless involves a particular schedule and reward structure, and a small experimental context cannot establish that all familiar routines show the same transition. Subsequent failures to induce habits with several human procedures require the observation to be considered part of an accumulating, heterogeneous evidence base. de Wit et al. (2018)

A habit signal in posterior putamen must also be distinguished from motor sequencing and practice-related skill acquisition. Brain responses can change because movements become more predictable, more efficient, or more strongly associated with a cue even when outcome evaluation remains possible. The decisive behavioral question is whether the learned response is controlled by present outcome value or contingency, not whether sensorimotor activation rises with practice. Cross-species comparison is strongest when the same behavioral criteria are preserved and when possible performance impairments are controlled. Balleine & O’Doherty (2010) articulates this challenge and situates striatal regions as components of parallel, interacting loops rather than isolated storage locations for entire habits.

Orbitofrontal Circuits and Reversible Control

An action can return to goal-directed control after previously being expressed habitually, suggesting that control strategies remain dynamically available. Gremel & Costa (2013) used a within-subject mouse task in which the same animals shifted across conditions favoring goal-directed and habitual responding. Orbitofrontal and striatal activity tracked these shifts, supporting flexible recruitment of control systems within an individual. The within-subject design reduces differences between animals and demonstrates that training history need not irreversibly determine behavioral organization. It also challenges a simple notion of habit as permanent replacement of earlier action knowledge. Instead, learned associations and their expression depend on current conditions and network state.

Pathway-specific experiments deepen the causal account. Gremel et al. (2016) found that manipulation of orbitofrontal projections to dorsal striatum and cannabinoid type 1 receptor signaling affected the shift toward habitual instrumental action in mice. The experimental result supports a mechanism in which modulation of a goal-directed circuit changes the balance between action-control strategies. Its interpretation is constrained by species, circuitry targeted, treatment timing, and the operational definition of habit in the paradigm. Molecular and optogenetic interventions provide evidence about necessity or modulation of the tested pathway; they do not imply a clinically established pharmacological strategy for habit modification in humans.

In another rat preparation, transient inactivation of infralimbic prefrontal cortex reinstated goal-directed sensitivity after overtraining had rendered behavior comparatively insensitive to outcome value. Coutureau & Killcross (2003) thereby implicated a cortical contribution to the expression of established response tendencies. Together with orbitostriatal studies, this suggests that switching can depend on suppressing or enabling competing control routes. The relevant effect is an experimental change in performance after an intervention, and the general network mechanism must be evaluated through independent replication and models that differentiate learning from expression. These findings motivate a view of habit control as a dynamic functional state, with value-based knowledge potentially retained even when a habitual response dominates.

Computational Reinforcement-Learning Accounts

Computational reinforcement learning provides explicit hypotheses about what information can guide action. A model-based learner represents states, action-dependent transitions, and outcome values, using this representation to estimate consequences before choosing. A model-free learner updates action values from realized rewards or prediction errors, without requiring a prospective search over a current transition model at each choice. These are algorithmic definitions. In animal learning, the goal-directed and habitual distinction is usually operationalized through sensitivity to devaluation and contingency, whereas model-based and model-free parameters are inferred by fitting a particular computational task. The pairs are related but not strictly interchangeable. Daw et al. (2005) Doll et al. (2012)

Model-free value updating can support efficient behavior after extensive feedback, but a cached value is not identical to a directly stored stimulus–response association in every formal model. An agent can update an action value in light of received rewards while retaining a relatively fixed policy between updates. A model-based agent can also produce repetitive behavior when the same action remains optimal. Thus, apparent routine action is compatible with either architecture; the identifying tests must place the architectures in situations where their predictions diverge. Daw et al. (2011) used a multistage human decision task to distinguish transition-sensitive choice from reward-based repetition and reported both influences on behavior and striatal prediction-error signals.

The distinction is most informative when the computational model makes falsifiable trial-level predictions. A model should specify its state representation, action set, transition probabilities, reward updating, exploration, choice noise, and assumed working-memory or search costs. Fitted parameters such as a weighting of model-based versus model-free values depend on that specification. A task may misclassify an agent if it learns hidden task structure, performs rapid inference, or uses an unmodeled strategy. Doll et al. (2012) notes the breadth of model-based influences, while de Wit & Dickinson (2009) connects formal approaches to associative action theory. Computational explanation gains credibility through simulation, parameter recovery, posterior predictive checks, and direct comparison with experimental interventions.

Arbitration, Reliability, and Resource Allocation

When multiple control policies are available, a further problem arises: which policy should determine an action now? Daw et al. (2005) proposed an uncertainty-sensitive competition in which the computational advantages of prospective model-based evaluation are balanced against its demands and the reliability of alternative cached estimates. This hypothesis explains why flexible control could be recruited selectively under environmental change and why practiced responses can dominate under stable conditions. The proposed arbitration mechanism is a model of behavioral allocation rather than an observed single neural switch. Its adequacy depends on whether predicted changes in control follow measured changes in uncertainty, cost, and reliability across tasks.

Human imaging provides tests of arbitration accounts. Lee et al. (2014) reported signals associated with estimates of the reliability of model-based and model-free predictions in lateral prefrontal regions, together with changes in functional connectivity that covaried with inferred control allocation. This result links explicit computation to neural correlates, but relies on fitted latent variables and assumptions about the underlying task. Reliability estimates, uncertainty measures, and model comparison are partly constructed from the same behavioral observations, creating potential identification problems. Alternative models can often reproduce similar choices by changing learning rates, exploration, or working-memory strategies. Independent perturbations are needed to distinguish a hypothesized arbitration computation from its behavioral correlates.

A balanced synthesis treats action control as conditional coordination among representational and procedural resources. In one situation, environmental stability favors a readily retrievable response; in another, unexpected reward loss increases the value of reconstructing action consequences. The empirical challenge is to establish where this allocation occurs, how rapidly it adapts, and whether previously learned habits can remain active without controlling overt performance. Hardwick et al. (2019) supports a response-competition interpretation at short timescales, while Gremel & Costa (2013) provides causal and recording evidence for within-subject switching in mice. Neither study alone determines a unique domain-general arbitration algorithm. Theory comparison should preserve this distinction between observed flexibility and inferred coordination rules.

Experimental Human Action-Selection Tasks

Human instrumental-learning tasks attempt to dissociate action–outcome control from stimulus-linked responding while preserving interpretable choice. Some train responses for distinct visual or food rewards and then change the value of one outcome; others require participants to avoid formerly rewarded outcomes after devaluation. A valid contrast depends on participants having learned the response–outcome pairings and understood the test condition, while the test itself must limit new reward feedback. de Wit et al. (2009) combined an outcome-sensitive task with neuroimaging to investigate differences in ventromedial prefrontal engagement. Such evidence reveals links between human task performance and neural signals but is conditioned by task instructions, awareness, learning criterion, and stimulus design.

The slips-of-action paradigm tests whether participants can suppress a learned response when its outcome has been devalued while continuing to respond for valued outcomes. Failure may index habitual tendency, but several other processes can cause slips: inaccurate memory, insufficient attention to the devaluation instruction, response conflict, or general inhibition problems. Conversely, perfect performance may arise from strategically slow responding rather than the absence of an acquired habit. Researchers should include outcome-knowledge tests, instrumental learning checks, training-dose comparisons, and time-pressured conditions. The strongest interpretations combine these indicators rather than treating a single proportion of slips as a pure trait measure. de Wit et al. (2018) Linnebank et al. (2018)

Researchers can also measure goal-directed response switching after learning that a formerly reinforced response is now disadvantageous. Luque et al. (2020) proposed reaction-time switch costs as an additional marker, showing variations with practice, reward magnitude, and time pressure. Hardwick et al. (2019) similarly showed that newly instructed action choices could coexist with rapidly prepared old responses. These findings expand the measurement portfolio beyond binary devaluation accuracy and reveal that response preparation is temporally structured. They do not automatically equate switch costs with stimulus–response learning: general task-set inertia, motor interference, and remapping difficulty remain viable alternative explanations requiring controlled comparison. A comprehensive assay should estimate selection and execution dynamics separately.

Why Human Overtraining Results Conflict

One major scientific controversy is whether extended training reliably produces the classical outcome-insensitivity observed in several rodent paradigms. Tricomi et al. (2009) reported reduced devaluation sensitivity after substantial human free-operant practice, with corresponding changes in posterior striatal cue responses. Yet de Wit et al. (2018) described five experiments that failed to achieve the anticipated shift from goal-directed to habitual responding under a range of induction conditions. The conflict is consequential because a theory predicting a general training-induced transition must confront failures as well as early demonstrations. Neither a positive study nor a series of null results alone fixes the boundary conditions without examining manipulation strength, measurement range, and alternative explanations.

Different studies vary in how responses are trained, which outcomes are used, what participants are told, and whether competing strategies remain available at test. Participants may create explicit rules for simple abstract tasks and use them even after extensive repetition. When a reward is meaningful or its identity is obvious, deliberate choice can dominate; when a devaluation intervention is weak, persistence may reflect inadequate updating rather than a habit. Training quantity is also not equivalent across procedures because trial spacing, response-outcome information, and context stability differ. de Wit et al. (2018) highlights these challenges, while Hardwick et al. (2019) demonstrates that reducing response preparation time can expose previously learned response tendencies that ordinary final-choice probes miss.

The appropriate response to discrepant evidence is a more discriminating research program. A new experiment should preregister the primary devaluation contrast and interaction with training dose, measure successful outcome revaluation independently, and test response-time distributions as well as error rates. Positive controls can verify that the task can detect outcome-sensitive responding before habit induction. Reliability across testing sessions and generalization to altered contexts should be assessed rather than assumed. The result may be that habits appear readily in some tasks, remain latent in others, and are difficult to distinguish from alternative mechanisms in a third class. Such a conditional account is scientifically stronger than treating every repetition effect as evidence for one universal habit transition. Luque et al. (2020) Wood & Rünger (2016)

Everyday Habit Automaticity and Longitudinal Observation

Naturalistic habit research often defines habit through repeated performance of a behavior in a recurring context and the growth of subjective automaticity. Lally et al. (2010) prospectively followed participants who selected simple daily activities and recorded performance and automaticity across weeks. Their fitted developmental curves supported gradual, often nonlinear increases, with considerable variation among participants and behaviors. This evidence is valuable because it captures behavior across ordinary settings rather than a single laboratory training session. Its principal outcome, however, was reported automaticity, not insensitivity to controlled outcome devaluation. The study therefore informs the acquisition of routine phenomenology and context consistency without establishing the full classical instrumental mechanism in each participant.

The Self-Report Habit Index was designed to assess features associated with repeated behavior, including automaticity and experiential regularity. Verplanken & Orbell (2003) developed and tested this instrument, while Gardner et al. (2012) evaluated an automaticity-oriented subscale. These measures support efficient longitudinal investigation and can distinguish frequently performed activities that feel effortful from those experienced as more automatic. They remain dependent on recall, introspection, normative expectations, and the meaning assigned to expressions such as doing something without thinking. A high score cannot function as direct proof of a specific neural circuit or a stimulus–response association. Combining self-report with unobtrusive behavioral sampling and controlled policy-change tests yields more interpretable evidence.

A useful integrative design separates three processes: recurrence of the context, performance of the behavior, and automaticity experienced or inferred by the actor. Stable contexts may increase opportunity for repetition, repetition may alter retrieval and preparation, and successful performance may reinforce beliefs that the activity is habitual. Wood & Rünger (2016) emphasizes how goal pursuit and habit can facilitate one another, while Linnebank et al. (2018) examined relationships between experimental habits and real-life routine measures. Correlation across the two domains is not guaranteed because laboratory choice and everyday automaticity are different operational constructs. Their comparison should estimate shared and unique variance and investigate whether a common mechanism predicts longitudinal change under context intervention.

Developmental, Contextual, and Individual Variation

Repeated actions are embedded in developmental histories and differing environments. Children and adults have different opportunities, working-memory resources, motivational priorities, and routines, but these differences do not license a simple monotonic claim that one age group is globally more habitual. The experimental literature in this bibliography establishes more directly the roles of reinforcement history, available response time, outcome identity, and contextual recurrence. A developmental inference requires age-comparable tasks and evidence that differences in devaluation sensitivity are not produced by language comprehension, memory, physical skill, or unequal value manipulation. Principles of construct validity and task comparability matter whenever age differences are interpreted as differences in underlying behavioral control. de Wit & Dickinson (2009) Wood & Rünger (2016)

Context stability can shape whether a practiced response is activated and whether altered intentions compete successfully. The same cue can mean different things across locations, social conditions, and time periods. Habit transfer therefore depends on the overlap of learned representations, not merely on counting past repetitions. In health behavior, consistent implementation contexts can support regular initiation, but changes in access, incentives, and environmental affordances also determine whether a response occurs. Lally et al. (2010) provides prospective evidence about repetition in stable settings, while Wood & Rünger (2016) frames context responsiveness as a key property of habits. Experimental transportability requires specifying which environmental features are causally relevant to the learned response.

Individual differences are particularly difficult to infer from a single laboratory session. Between-person variation in apparently habitual choices can reflect training history, reward preference, response caution, fatigue, and the reliability of outcome learning. Some experimental tasks measure a balance of processes at one moment rather than a stable general tendency. Linnebank et al. (2018) investigated the relation of laboratory markers to real-life habit formation, illustrating the importance of validating cross-context trait claims. A rigorous individual-differences program requires test–retest assessment, adequate measurement reliability, multiple tasks, and correction for task-specific variance. It should also examine whether similar behavior arises from distinct configurations of knowledge, response preparation, and situational opportunity.

Stress, Motivational State, and Control Flexibility

Acute stress is a candidate modifier of action control because it can change arousal, attention, memory retrieval, and the relative engagement of control circuits. Schwabe & Wolf (2009) reported experimental evidence that stress could promote habitual responding in a human instrumental task. Subsequent pharmacological and neuroendocrine work investigated contributions of glucocorticoid and noradrenergic activity. Schwabe et al. (2010) reported that concurrent activation of these systems influenced the balance of instrumental control under its tested conditions. These findings support a mechanistic research hypothesis about stress-dependent changes in action selection, while the size and generality of the effect vary with intervention, timing, population, and behavioral assay.

Stress may impair flexible use of changed outcome value without necessarily strengthening a newly acquired stimulus–response association. An individual under time pressure might rely more on an existing response, experience reduced working-memory capacity for counterfactual evaluation, or fail to encode the new devaluation instruction. Those possibilities are not interchangeable, and they require experiments separating acquisition from expression. Hormonal measures alone cannot prove a particular action policy; response tests and manipulation checks remain decisive. Research should therefore distinguish the effect of stress before training, immediately before valuation, and at response selection. Schwabe et al. (2010) Hardwick et al. (2019)

Longer-term application requires care because laboratory stressors and chronic real-world adversity differ in duration, controllability, and consequences. A person may pursue valued goals with considerable persistence while under stress, and some repeated behaviors may remain sensitive to new incentives. Broad claims that stress simply eliminates deliberate action obscure these conditions. The useful scientific formulation identifies an experimentally tested shift in the relative expression of outcome-sensitive and practiced response patterns under specific stress manipulations. Wood & Rünger (2016) connects habit research to applied behavioral change, and Schwabe & Wolf (2009) supplies a clear example of a laboratory perturbation. Further research should combine physiological measures, explicit knowledge checks, and repeated behavior sampling.

Compulsivity, Addiction, and Clinical Boundaries

Clinical research has investigated whether impaired goal-directed control contributes to repetitive behaviors, including obsessive-compulsive symptoms and substance-related behavior. Gillan et al. (2011) compared individuals with obsessive-compulsive disorder and controls using instrumental-learning methods and reported differences consistent with altered balance between goal-directed choice and habitual responding. This finding provides a translational hypothesis, not a diagnostic assay: a single laboratory measure cannot establish obsessive-compulsive disorder or identify the cause of an individual's symptoms. Clinical repetition encompasses intrusive thoughts, distress, avoidance, reinforcement, uncertainty, and functional impairment, each of which may interact with action selection in different ways. A sound analysis preserves the distinction between experimental mechanism and clinical syndrome.

Dimensional evidence adds another level of inquiry. Gillan et al. (2016) linked reduced goal-directed control to a symptom dimension involving compulsive behavior and intrusive thought in two general-population samples. This approach avoids assuming that diagnostic groups map onto discrete learning mechanisms and points toward quantitative variation across people. Nevertheless, association does not demonstrate that diminished goal-directed control caused the symptoms, that all affected individuals possess stronger stimulus–response habits, or that a manipulation improving one laboratory score will improve clinical functioning. Sampling, symptom measurement, task reliability, confounding, and direction of effects remain decisive. The results are best interpreted as support for a research relationship that needs longitudinal and intervention-based tests.

Models of addiction have proposed that repeated reinforcement and exposure to drug-related cues can shift the determinants of seeking and consumption, with habitual and compulsive processes becoming increasingly prominent. Everitt & Robbins (2016) critically reviewed this progression and its neural circuitry, including interactions between Pavlovian stimuli, instrumental actions, and regulatory control. The progression is a theoretical framework informed by preclinical and human evidence; it is not an invariant sequence through which every person passes. Substance-related behaviors have pharmacological, motivational, social, developmental, and treatment-specific determinants. The present dossier addresses general action-control mechanisms, while diagnosis, exposure procedures, and pharmacological interventions belong to dedicated clinical research and qualified care.

Behavior Change and Reorganization of Habits

Behavior change can be understood as modification of the relations among cues, response opportunities, goals, and expected outcomes. For constructing a beneficial routine, selecting a feasible action and repeating it in a recognizable context may increase its ease of initiation over time. Lally et al. (2010) documented the development of self-reported automaticity during daily repetition, but the considerable heterogeneity of trajectories precludes one universal timetable. For replacing an unwanted behavior, changing the environment can reduce the frequency of cues that initiate the old response, while a specifically practiced alternative can become available in the same situation. These are testable applications of context-dependent learning rather than guarantees of individual success.

Outcome revaluation and contingency learning offer complementary routes to flexibility. If the expected consequence has become less desirable, the actor may benefit from making this change salient at the moment of choice. If the old action no longer effectively produces its desired consequence, explicit comparison of available alternatives can support adaptation. Yet awareness of changed goals may coexist with a prepared response, especially under time pressure. Hardwick et al. (2019) shows why interventions should measure both planned decisions and rapid action expression. The mechanism of a successful intervention must be established by observing changes in behavior under controlled cue, value, and opportunity conditions rather than assuming a cognitive intention alone is sufficient.

Practical strategies require scientific evaluation of adherence, generalization, unintended substitution, and durability. A routine that appears stable during a brief intervention may disappear when context changes; a new goal may work in a laboratory but not in a demanding daily environment. Prospective behavioral records, repeated automaticity measures, and experimentally manipulated cues can help identify what changed. Gardner et al. (2012) supports parsimonious assessment of experienced automaticity, and Wood & Rünger (2016) connects habit research to behavior-change design. These findings justify an experimental program involving cue restructuring and alternative-response rehearsal, while clinical treatment decisions remain subject to disorder-specific evidence and professional judgment.

Measurement Validity, Reliability, and Causal Identification

Goal-directed and habitual control are theoretical constructs inferred from behavioral contrasts, fitted models, and neural measures. Construct validity depends on whether each indicator behaves as its theoretical definition predicts and discriminates the construct from memory, movement speed, reward knowledge, inhibition, and explicit strategy use. A devaluation assay requires evidence that the outcome lost value; a degradation assay requires evidence that the contingency changed; an automaticity questionnaire requires evidence that its items capture an experienced process rather than merely frequency. Verplanken & Orbell (2003) Gardner et al. (2012) provide measurement-development work, while de Wit et al. (2018) illustrates how experimental induction can fail to produce the expected construct contrast.

Reliability and causal identification are distinct. A highly reliable switch cost can consistently measure processes other than habit, whereas a theoretically specific measure may be noisy if it relies on a small number of critical choices. Designs should report learning criteria, exclusion rules, manipulation checks, response opportunities, effect estimates, uncertainty intervals, and test–retest information. The target outcome should be specified before analysis, especially when multiple possible indices of slips, response rates, reaction times, or computational parameters are available. Luque et al. (2020) advances a response-switch measure but also motivates direct tests against motoric and executive-control alternatives. Repeated individual observations can improve precision if dependence across trials is modeled appropriately.

Neural inference also demands care. A BOLD correlate of a fitted learning parameter does not demonstrate that the brain region implements a uniquely specified algorithm, and an intervention's effect may propagate through connected networks. Lesions, pharmacological manipulations, and pathway-specific techniques offer different kinds of causal information and different potential confounds. Yin et al. (2005) supports a causal contribution of posterior dorsomedial striatum under a specified rat paradigm; Gremel et al. (2016) tests pathway involvement in mice; Lee et al. (2014) identifies human imaging correlates of computational arbitration. The strongest mechanistic account triangulates these results without treating anatomical correspondence, behavioral fit, and clinical causation as the same claim.

Competing Interpretations and Open Research Problems

The classical dual-process distinction has substantial explanatory value because it creates clear predictions about how action selection should respond to value and contingency. Its limitations arise when observed behavior can be generated by several architectures or when available measures fail to isolate the process of interest. A model-based learner may repeat an action because it remains optimal; a model-free learner may respond adaptively after new reward experience; explicit instructions can masquerade as devaluation sensitivity. Daw et al. (2011) revealed combined model-based and model-free influences in both choice and striatal signals, complicating a rigid anatomical or computational separation. An adequate theory should specify the causal manipulation that would discriminate it from such alternatives.

One open problem concerns the timing of control. The completed response is the endpoint of perception, retrieval, preparation, inhibition, and execution, but habits may influence intermediate stages without determining the final action. Hardwick et al. (2019) supplies evidence for latent practiced response preparation that can be overridden given time. Another problem concerns the representational unit: does a habit attach to a visual cue, a context, an action chunk, a task state, or an inferred situation? The answer may differ across tasks and training histories. More temporally resolved measurement, preregistered behavioral perturbations, and computational modeling with explicit recovery tests can address these questions.

Generalization remains a third unresolved problem. Animal learning studies offer strong mechanistic interventions, human instrumental tasks measure narrow experimental choices, and real-world automaticity surveys track extended activity in natural settings. Their common language is useful, but validity does not flow automatically from one domain to another. Lally et al. (2010) measures longitudinal automaticity, de Wit et al. (2018) documents challenges of human laboratory induction, and Gillan et al. (2016) relates task performance to symptom dimensions. A coherent research program must establish measurement invariance, predictive validity beyond training, and transfer across context changes. Independent replications and transparent reporting are especially important where the evidence base contains demonstrably conflicting task outcomes.

Artificial Era and Cross-System Action Control

Artificial systems make the architecture of action control explicit enough to compare algorithms under controlled changes in reward, transition structure, memory, and computation. An artificial agent may evaluate candidate actions using a transition model, retrieve previously successful responses, or arbitrate between these procedures according to resource costs. These mechanisms are formal functional comparators for model-based and model-free accounts in psychological research. Daw et al. (2005) Lee et al. (2014) offer frameworks in which controller selection is itself a computational problem. The corresponding psychological question remains empirical: which aspects of biological learning are reproduced by an algorithm, and which require additional assumptions about perception, valuation, embodiment, and development?

A software policy that repeats a response under a particular input is not automatically the same scientific object as a learned biological habit. For comparison, an experiment must specify the artificial system's input distribution, training history, action opportunities, value update, memory architecture, and behavior after the outcome and contingency interventions. Such tests can reveal whether the system behaves as if it retained prospective outcome sensitivity or relied on a previously established policy. These are observations of functional organization, and their interpretation depends on accessible mechanisms and the test design. The distinction prevents anthropomorphic classification from replacing operational analysis, while retaining artificial action systems as a meaningful research domain within the Artificial Era.

Human–Artificial partnerships introduce further research questions because people may delegate choices, follow repeated recommendations, or develop context-dependent practices around a tool. A human's habitual use of an interface, an algorithm's recurrent policy, and the jointly produced behavior of a human–artificial system are different units of analysis. Investigators should identify whose action is measured, which outcomes each party can observe, and how repeated interaction reshapes expectations and response opportunities. The theoretical connection to distributed action control is promising, but current instrumental-learning evidence does not establish a general equivalence between artificial policy optimization and human psychological habit. Cross-system research should therefore rely on matched interventions, transparent task descriptions, and convergent behavioral evidence. Wood & Rünger (2016) Balleine & O’Doherty (2010)

Recent Preregistered Replications and Convergent-Validity Tests

Research published in 2025–2026 directly tests whether established habit assays identify the processes they purport to measure. Molinero et al. (2025) conducted three preregistered human experiments involving three days of instrumental training, contrasting mappings practiced at different frequencies and evaluating behavior after action–outcome contingencies were degraded. Overtrained responding was not more habitual than the comparison condition on the tested measure. Participants frequently reduced responding when the contingency changed, while apparently habit-like actions could correspond to incomplete recognition of the manipulation. The result challenges the assumption that adding training necessarily generates observable contingency-insensitive control and illustrates why contingency perception is part of the validity argument, not an optional post hoc description.

A separate cross-task psychometric investigation Martínez-López et al. (2026) evaluated two time-pressure-based habit measures after four days of training. Responses in a forced-response paradigm and reaction-time switch costs in an outcome-devaluation paradigm each showed favorable within-paradigm reliability, but the individual differences on the two measures were not correlated. This combination is theoretically important. A response can be stable across repeated tests and still fail to converge with a different instrument intended to measure the same latent process. Apparent habituation in one task may reflect response preparation and reversal difficulty, whereas another index reflects slowed adjustment after a reward reassignment. The study strengthens the requirement to specify which component of action control a measure identifies and to validate intertask convergence empirically.

Most recently, Vázquez-Millán et al. (2026) reported a preregistered conceptual replication of a human free-operant task using an improved outcome-devaluation protocol and reanalysis of earlier datasets. The researchers observed no straightforward overtraining contrast and found that the participants appearing most habitual often showed the least effective outcome devaluation. When the reward-value manipulation is weak, continued action can remain consistent with the participant's current value representation. These findings establish a specific methodological warning: behavior can appear insensitive to an experimenter's intended value change because the participant did not experience that change. Convergent across Molinero et al. (2025), Martínez-López et al. (2026), and Vázquez-Millán et al. (2026) is the need for individual-level manipulation checks, multiple discriminating assays, temporal measures, and careful calibration of what the label habit means in each experiment.

Conclusion

The evidence establishes goal-directed action and habitual control as indispensable operational categories for the science of learned behavior. Outcome devaluation tests whether current incentive value governs an acquired action; contingency degradation tests whether the actor remains sensitive to the action's contribution to producing its consequence. These manipulations distinguish important aspects of control that response frequency and proficiency cannot establish. Classic instrumental experiments and corticostriatal causal interventions provide especially strong support for the behavioral distinction in controlled animal preparations. Human studies establish flexible outcome-based action and provide demonstrations of practiced response tendencies, together with unresolved conflicts over the reliability and conditions of laboratory habit induction. Adams & Dickinson (1981) Yin et al. (2004) de Wit et al. (2018)

The scientific synthesis favors an interactional account of action control in which knowledge of consequences, cached response tendencies, available time, contextual cues, and motivational state jointly determine performance. Neural data implicate distributed corticostriatal circuitry, with distinct but interacting contributions of dorsomedial, dorsolateral, and frontal systems. Computational studies make possible controllers and their arbitration explicit, while their fitted parameters remain dependent on task and model assumptions. Everyday longitudinal studies show developing automaticity under repeated contextual performance, although such automaticity is not itself a direct devaluation assay. The strongest conclusions follow from connecting these evidence streams according to the specific questions each can answer. Gremel & Costa (2013) Daw et al. (2011) Lally et al. (2010)

Further progress requires improved manipulation checks, reliable experimental measures, adequate training-dose contrasts, explicit tests of response-preparation timing, and preregistered replication across populations and environments. Causal neural interventions should be linked to behaviorally identified functions, while computational modeling should include parameter recovery and alternative-strategy comparison. Clinical and Artificial Era applications demand independently validated constructs and careful transfer of mechanism claims. The final scientific position is that goal-directedness and habit describe experimentally distinguishable modes of behavioral organization whose balance depends on learning, valuation, context, neural coordination, and conditions of expression. Their investigation remains central to explaining how action becomes efficient while retaining the potential for flexible revision. Balleine & O’Doherty (2010) Hardwick et al. (2019) Wood & Rünger (2016)

References

  • Adams, C. D., & Dickinson, A. (1981). Instrumental responding following reinforcer devaluation. The Quarterly Journal of Experimental Psychology: Section B, 33(2), 109–121. https://doi.org/10.1080/14640748108400816
  • Balleine, B. W., & Dickinson, A. (1998). Goal-directed instrumental action: Contingency and incentive learning and their cortical substrates. Neuropharmacology, 37(4–5), 407–419. https://doi.org/10.1016/S0028-3908(98)00033-1
  • Balleine, B. W., & O’Doherty, J. P. (2010). Human and rodent homologies in action control: Corticostriatal determinants of goal-directed and habitual action. Neuropsychopharmacology, 35(1), 48–69. https://doi.org/10.1038/npp.2009.131
  • Coutureau, E., & Killcross, S. (2003). Inactivation of the infralimbic prefrontal cortex reinstates goal-directed responding in overtrained rats. Behavioural Brain Research, 146(1–2), 167–174. https://doi.org/10.1016/j.bbr.2003.09.025
  • 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
  • Daw, N. D., Gershman, S. J., Seymour, B., Dayan, P., & Dolan, R. J. (2011). Model-based influences on humans’ choices and striatal prediction errors. Neuron, 69(6), 1204–1215. https://doi.org/10.1016/j.neuron.2011.02.027
  • de Wit, S., & Dickinson, A. (2009). Associative theories of goal-directed behaviour: A case for animal–human translational models. Psychological Research, 73(4), 463–476. https://doi.org/10.1007/s00426-009-0230-6
  • de Wit, S., Corlett, P. R., Aitken, M. R., Dickinson, A., & Fletcher, P. C. (2009). Differential engagement of the ventromedial prefrontal cortex by goal-directed and habitual behavior toward food pictures in humans. The Journal of Neuroscience, 29(36), 11330–11338. https://doi.org/10.1523/JNEUROSCI.1639-09.2009
  • de Wit, S., Kindt, M., Knot, S. L., Verhoeven, A. A. C., Robbins, T. W., Gasull-Camos, J., Evans, M., Mirza, H., & Gillan, C. M. (2018). Shifting the balance between goals and habits: Five failures in experimental habit induction. Journal of Experimental Psychology: General, 147(7), 1043–1065. https://doi.org/10.1037/xge0000402
  • Doll, B. B., Simon, D. A., & Daw, N. D. (2012). The ubiquity of model-based reinforcement learning. Current Opinion in Neurobiology, 22(6), 1075–1081. https://doi.org/10.1016/j.conb.2012.08.003
  • Everitt, B. J., & Robbins, T. W. (2016). Drug addiction: Updating actions to habits to compulsions ten years on. Annual Review of Psychology, 67, 23–50. https://doi.org/10.1146/annurev-psych-122414-033457
  • Gardner, B., Abraham, C., Lally, P., & de Bruijn, G.-J. (2012). Towards parsimony in habit measurement: Testing the convergent and predictive validity of an automaticity subscale of the Self-Report Habit Index. International Journal of Behavioral Nutrition and Physical Activity, 9, Article 102. https://doi.org/10.1186/1479-5868-9-102
  • Gillan, C. M., Papmeyer, M., Morein-Zamir, S., Sahakian, B. J., Fineberg, N. A., Robbins, T. W., & de Wit, S. (2011). Disruption in the balance between goal-directed behavior and habit learning in obsessive-compulsive disorder. American Journal of Psychiatry, 168(7), 718–726. https://doi.org/10.1176/appi.ajp.2011.10071062
  • Gillan, C. M., Kosinski, M., Whelan, R., Phelps, E. A., & Daw, N. D. (2016). Characterizing a psychiatric symptom dimension related to deficits in goal-directed control. eLife, 5, Article e11305. https://doi.org/10.7554/eLife.11305
  • Gremel, C. M., & Costa, R. M. (2013). Orbitofrontal and striatal circuits dynamically encode the shift between goal-directed and habitual actions. Nature Communications, 4, Article 2264. https://doi.org/10.1038/ncomms3264
  • Gremel, C. M., Chancey, J. H., Atwood, B. K., Luo, G., Neve, R., Ramakrishnan, C., Deisseroth, K., Lovinger, D. M., & Costa, R. M. (2016). Endocannabinoid modulation of orbitostriatal circuits gates habit formation. Neuron, 90(6), 1312–1324. https://doi.org/10.1016/j.neuron.2016.04.043
  • Hardwick, R. M., Forrence, A. D., Krakauer, J. W., & Haith, A. M. (2019). Time-dependent competition between goal-directed and habitual response preparation. Nature Human Behaviour, 3(12), 1252–1262. https://doi.org/10.1038/s41562-019-0725-0
  • Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., & Wardle, J. (2010). How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology, 40(6), 998–1009. https://doi.org/10.1002/ejsp.674
  • Lee, S. W., Shimojo, S., & O’Doherty, J. P. (2014). Neural computations underlying arbitration between model-based and model-free learning. Neuron, 81(3), 687–699. https://doi.org/10.1016/j.neuron.2013.11.028
  • Linnebank, F. E., Kindt, M., & de Wit, S. (2018). Investigating the balance between goal-directed and habitual control in experimental and real-life settings. Learning & Behavior, 46(3), 306–319. https://doi.org/10.3758/s13420-018-0313-6
  • Luque, D., Molinero, S., Watson, P., López, F. J., & Le Pelley, M. E. (2020). Measuring habit formation through goal-directed response switching. Journal of Experimental Psychology: General, 149(8), 1449–1459. https://doi.org/10.1037/xge0000722
  • Martínez-López, P., Vázquez-Millán, A., Garre-Frutos, F., & Luque, D. (2026). Assessing the validity evidence for habit measures based on time pressure. Behavior Research Methods, 58, Article 7. https://doi.org/10.3758/s13428-025-02865-2
  • Molinero, S., Martínez-López, P., Morís, J., Quintero, M. J., Cobos, P. L., López, F. J., & Luque, D. (2025). The degraded contingency test fails to detect habit induction in humans. PLOS ONE, 20(10), Article e0334087. https://doi.org/10.1371/journal.pone.0334087
  • Schwabe, L., & Wolf, O. T. (2009). Stress prompts habit behavior in humans. The Journal of Neuroscience, 29(22), 7191–7198. https://doi.org/10.1523/JNEUROSCI.0979-09.2009
  • Schwabe, L., Tegenthoff, M., Höffken, O., & Wolf, O. T. (2010). Concurrent glucocorticoid and noradrenergic activity shifts instrumental behavior from goal-directed to habitual control. The Journal of Neuroscience, 30(24), 8190–8196. https://doi.org/10.1523/JNEUROSCI.0734-10.2010
  • Tricomi, E., Balleine, B. W., & O’Doherty, J. P. (2009). A specific role for posterior dorsolateral striatum in human habit learning. European Journal of Neuroscience, 29(11), 2225–2232. https://doi.org/10.1111/j.1460-9568.2009.06796.x
  • Valentin, V. V., Dickinson, A., & O’Doherty, J. P. (2007). Determining the neural substrates of goal-directed learning in the human brain. The Journal of Neuroscience, 27(15), 4019–4026. https://doi.org/10.1523/JNEUROSCI.0564-07.2007
  • Vázquez-Millán, A., Martínez-López, P., Rueda, M., León, J. J., & Luque, D. (2026). The evaluation of devaluation: Deficient outcome devaluation leads to wrongly considering goal-directed actions as habits. Behavior Research Methods, 58, Article 224. https://doi.org/10.3758/s13428-026-03099-6
  • Verplanken, B., & Orbell, S. (2003). Reflections on past behavior: A self-report index of habit strength. Journal of Applied Social Psychology, 33(6), 1313–1330. https://doi.org/10.1111/j.1559-1816.2003.tb01951.x
  • Wood, W., & Rünger, D. (2016). Psychology of habit. Annual Review of Psychology, 67, 289–314. https://doi.org/10.1146/annurev-psych-122414-033417
  • Yin, H. H., Knowlton, B. J., & Balleine, B. W. (2004). Lesions of dorsolateral striatum preserve outcome expectancy but disrupt habit formation in instrumental learning. European Journal of Neuroscience, 19(1), 181–189. https://doi.org/10.1111/j.1460-9568.2004.03095.x
  • Yin, H. H., Ostlund, S. B., Knowlton, B. J., & Balleine, B. W. (2005). The role of the dorsomedial striatum in instrumental conditioning. European Journal of Neuroscience, 22(2), 513–523. https://doi.org/10.1111/j.1460-9568.2005.04218.x