RESEARCH DOSSIER
Pavlovian Conditioning and Contingency Learning
Abstract. Pavlovian conditioning and contingency learning concern how organisms acquire predictions about consequential events from signals, background event rates, temporal relations and prior learning. This fundamental critical narrative monograph reconstructs Pavlov's physiological research and contemporary experiments on contingency, cue competition, blocking, overshadowing, inhibitory learning, prediction error and conditioned response. It compares the Rescorla–Wagner model, attentional associability theories, temporal information models, configural learning, human expectancy and causal inference, and Bayesian and temporal-difference accounts. Primary evidence from comparative conditioning, human threat-learning paradigms, behavioral and neural studies is evaluated alongside contemporary 2025 work on prospective contingency and dopamine signals. Separate chapters address methods, reliability, computational identifiability, response expression, artificial learning systems, implications and unresolved scientific questions. The synthesis distinguishes experimentally arranged contingencies, inferred representations and measured conditioned behavior, and identifies conditions that separate competing theoretical explanations. The work is a critical narrative research synthesis of published literature and reports no original experiments, systematic-review protocol or meta-analysis.
Keywords: Pavlovian conditioning; classical conditioning; contingency learning; predictive learning; CS–US contingency; blocking; overshadowing; conditioned inhibition; Rescorla–Wagner model; prediction error; associability; temporal learning; cue competition; human expectancy; causal inference; dopamine; temporal-difference learning; Artificial Era
The Scientific Object of Predictive Learning
Pavlovian conditioning is the acquisition and use of relations among environmental events in which a conditioned stimulus (CS) conveys information about an unconditioned stimulus (US) or another consequential event. The observable conditioned response (CR) may involve approach, withdrawal, freezing, salivation, autonomic adjustment, or a learned expectancy, depending on the species, outcome, task, and motivational system. The research problem is to explain how the organism extracts predictive information from a sequence of events, how it represents the relation, and under which retrieval and performance conditions that information governs behavior. Pavlovian learning therefore concerns the architecture of expectation and action preparation, not merely repetition of an originally unconditioned movement Pavlov (1927) Rescorla (1988) Harris (2025).
The conditioned stimulus is initially specified by its experimental function, not by an intrinsic physical property; a light, odor, tone, visual pattern, or context can serve as a CS. The US is an event whose motivational or physiological effects have been established independently of the relevant training history. A CS becomes experimentally meaningful through its programmed relation to that event. The CR is not necessarily identical in form to the unconditioned response (UR): food-predicting cues can organize approach toward the cue or food magazine, while threatening events recruit species-characteristic defensive patterns. An informative account of conditioning distinguishes changes in acquired knowledge from the rules transforming that knowledge into performance Rescorla (1988) Bouton & Moody (2004).
This monograph addresses positive and negative contingency, intertrial background rates, cue competition, conditional and configural discrimination, timing, associability, outcome representation, human awareness, comparative causal inference, and neural prediction signals. Its evidentiary method is a critical narrative synthesis of published experiments, formal models, and methodological reviews, without claiming an exhaustive systematic search or quantitative meta-analysis. Extinction and renewal are considered where they reveal what has been learned, while the independent dossier on extinction retains the broader treatment of recovery phenomena; action–outcome learning and neural implementations also have separate research boundaries. The organizing question throughout is which experimental result discriminates among alternative accounts of a learned prediction Miller et al. (1995) Harris (2025).
Pavlov's Experimental Physiology and the Signal Relation
Ivan P. Pavlov's laboratory developed controlled preparations in which food-related salivary secretion could be measured in response to previously neutral signals. His 1927 English-language collection, translated by G. V. Anrep, described conditioned reflexes as an experimental method for investigating cerebral functions, including stimulus discrimination, generalization, inhibition, timing, and changing physiological states. It documented a broad research program involving researchers and assistants rather than a single anecdotal demonstration. The essential historical advance was the ability to impose precise relations between a stimulus and a biologically consequential event and trace the emergence of anticipatory responding across trials Pavlov (1927).
The historical term conditioned reflex reflected Pavlov's physiological explanatory commitments, but the behavior studied by his school was richer than a one-to-one transfer of a response from food to a bell. Organisms discriminated signals of different duration and configuration; responding depended on temporal proximity, reinforcement history, and the relations of stimuli within compounds. The experiments stimulated comparative learning research and later mathematical analyses because stimulus–outcome sequences could be systematically manipulated while behavioral output was observed. What survived most productively was the empirical paradigm of experimentally arranged signals and consequences, which supports multiple competing theories of psychological representation Pavlov (1927) Rescorla (1988).
A historical explanation should separate Pavlov's experimental observations from his era's cortical excitation–inhibition language. That vocabulary motivated important tests, but contemporary descriptions require independent neural evidence when attributing conditioning effects to specific circuits. The salivary preparation was especially useful for its observable output, yet its scope did not predetermine which mechanisms would explain freezing, reward approach, emotional valuation, or human causal judgment. Later studies expanded conditioning beyond a physiological reflex into a general experimental method for studying acquired environmental knowledge Rescorla (1988) Schultz et al. (1997) McNally et al. (2011).
Operational Definitions: Contiguity, Contingency, and Prediction
Temporal contiguity refers to proximity between an event designated CS and the outcome designated US. Contingency concerns how the occurrence or timing of the US differs when the CS is present or absent, or compared with an appropriate baseline process. In simple discrete-trial notation, ΔP = P(US|CS) − P(US|not CS) expresses a difference in conditional probability. A positive value indicates that the cue increases the probability of the outcome; a negative value indicates that its occurrence reduces that probability. The statistic is helpful only when the observation window, event boundaries, exposure duration, and baseline sampling convention are fully specified Rescorla (1968) Harris (2025).
Two experimental schedules may contain the same number of CS–US pairings yet support different predictive inferences. When unsignaled outcomes become frequent between trials, the informational advantage of the cue can diminish even though its own paired presentations remain intact. Conversely, a cue may predict an outcome effectively when outcome rate during the cue differs sharply from the background rate. These designs require controls for temporal spacing, cumulative exposure, sensitization, and unconditioned changes in response level. Rescorla's 1968 experiment made contingency central by manipulating shock probability in the presence and absence of a cue, a historical turning point in the empirical analysis of associative learning Rescorla (1968).
Prediction is both a theoretical construct and an experimental target. It can be assessed through anticipatory behavior, outcome-specific transfer, verbal expectancy ratings, and neural signals, but agreement among these measures is an empirical matter. An observed difference in conditioned responding may be consistent with a change in learning, a change in motivation, or a change in expression at the test. Explicitly separating acquisition, retention, retrieval, and performance protects the meaning of contingency experiments and defines which type of claim each measurement can establish Bouton & Moody (2004) Lonsdorf et al. (2017).
Experimental Control and the Paired–Unpaired Comparison
The basic acquisition design repeatedly presents a target cue and a consequential outcome, then measures a response during cue presentation relative to baseline or to a control cue. Its crucial controls concern the conditional distribution of the outcome, the total number of outcomes, temporal placement, and prior experiences of the stimulus. Simply contrasting paired trials with an empty control condition confounds cue learning with experience of the outcome. Unpaired and explicitly uncorrelated conditions can address some of these alternatives, although their interpretation also depends on whether an unpaired schedule inadvertently establishes a negative contingency Rescorla (1968) Harris (2025).
Experimental designs differ in whether they present a discrete trial, a continuous stream of events, or a temporally extended context. Trial counts obscure the time the participant spent waiting for outcomes; consequently trials with identical nominal reinforcement ratios may deliver different information about hazard or outcome rates. The learning curve should be analyzed with sufficient temporal resolution to distinguish initial acquisition, asymptotic performance, and transient changes after a schedule transition. Within-subject reversals and carefully counterbalanced controls may reveal rapid updating that between-group final scores conceal Gallistel & Gibbon (2000) Balsam & Gallistel (2009) Harris (2025).
Response selection matters. Freezing can conceal competing movements; a low skin-conductance response may reflect physiological nonresponsiveness rather than absence of threat expectancy; a high verbal probability judgment need not generate a matching autonomic signal. Modern human threat-learning guidelines therefore recommend reporting the stimulus schedule, reinforcement rates, the definition of nonresponders, calibration procedures, handling of zero responses, and whether dependent measures are interpreted separately or jointly. A conditioning study becomes informative when its design and measurement jointly exclude plausible alternatives Lonsdorf et al. (2017).
From Reflex Acquisition to Learning Relations Among Events
Robert Rescorla's later reconceptualization framed Pavlovian conditioning as learning relations among events that support a representation of the environment. The transition did not discard behavioral measurement; it enlarged the class of hypotheses tested by that measurement. If learning depends on which outcomes are expected, it should respond to manipulations of outcome identity, schedule, stimulus configuration, and expectations established by other cues. Such manipulations reveal why the description of a neutral stimulus acquiring a reflex-like response is incomplete as a scientific explanation Rescorla (1988).
The relational interpretation also changes how a conditioned response is analyzed. A target CS may acquire excitatory properties from a relation with food, but its expression can depend on motivational state, background context, competition from other signals, and the current inferred structure of the environment. Animals may behave as though multiple learned relations are available, with one relation controlling performance on a particular trial. Comparing outcome-specific transfer, compound-cue responding, and tests following separate outcome experience offers stronger evidence about representational content than measuring response magnitude alone Miller et al. (1995) Bouton & Moody (2004).
This shift provides the scientific foundation for the research program developed in the following chapters: associative-strength models specify quantitative rules for updating predictions; attentional models specify how stimulus processing itself changes; temporal models encode when events should occur; inferential models distinguish observed correlation from intervention; and neural studies seek physiological variables corresponding to these computational quantities. These approaches overlap empirically but do not become equivalent by sharing the word prediction. Their relative merit is established by successful risky predictions and well-controlled tests Rescorla & Wagner (1972) Mackintosh (1975) Gallistel & Gibbon (2000) Gershman (2015).
Blocking and the Information Added by a New Cue
Blocking is a central test of whether co-occurrence with an outcome is sufficient for associative acquisition. In the standard sequence, cue A first predicts the outcome; subsequently the compound AX predicts the same outcome. At test X alone typically supports less conditioned responding than a comparable novel cue trained in a compound without prior A reinforcement. Because X has been paired with the outcome during compound trials, the result identifies the informational contribution of a cue as an important determinant of learning. Prior expectation based on A alters what can be learned about X Rescorla & Wagner (1972) Miller et al. (1995).
Blocking is often interpreted as evidence that little prediction error remains when the outcome is already fully expected, but this interpretation is not unique. Attentional allocation may make X less effectively processed, and reasoning about whether an added cue makes a causal difference may constrain learning in some tasks. The strongest tests separate associative value, learned associability, cue salience, and a causal account of redundancy by using modified pretraining schedules, explicit additional-outcome information, and tests of retrospective change. Experimental work in rats demonstrates that forward blocking can be sensitive to constraints drawn from causal inference Beckers et al. (2006).
Blocking also depends on performance at test. A weak response to X does not alone disclose whether X was poorly encoded, encoded with low associative value, or encoded in a manner masked by comparator processes. Across laboratories and species, stimulus modality, compound perceptual organization, context, reinforcement density, and acquisition history can alter the outcome. The phenomenon is best treated as a family of informative cue-competition results rather than an infallible signature of one updating algorithm Miller et al. (1995) Pearce (1994).
Overshadowing, Cue Salience, and Compound Representations
Overshadowing occurs when conditioning to one element is diminished by concurrent training with another element, often one more salient or informative under the task conditions. The elements may compete for a limited common outcome prediction or differ in their probability of being attended to. Physical intensity is one determinant of salience but cannot serve as a universal scalar independent of the organism's sensory system and prior experience. Tests should manipulate both cue effectiveness and compound configuration to distinguish stimulus dominance from relative predictive validity Wagner et al. (1968) Mackintosh (1975).
A compound AX need not be represented as the independent sum of elemental features A and X. Animals can encode configurations, relations, or unique compound features, making the response to a component depend on the similarity between its training representation and its test representation. Pearce's 1994 connectionist configural account explains discrimination phenomena that challenge simple elemental theories. In evaluating overshadowing, researchers therefore ask whether limited learning to an element reflects allocation of associative strength, attention to the element, or generalization from the compound representation Pearce (1994).
The developmental and practical implications follow from the experimental logic: cue competition predicts that adding a highly salient signal during training can change which aspects of a situation later control behavior. That prediction requires a transfer test when each cue appears alone. In educational, human–machine, and clinical contexts, salient instructions, alarms, or interface indicators can plausibly compete with a less conspicuous environmental feature. Yet such applications remain hypotheses until studied in the relevant population and task, with controls for comprehension, explicit instruction and outcome utility Mackintosh (1975) Harris (2025).
The Rescorla–Wagner Error-Correction Model
The Rescorla–Wagner account represents the associative change to each cue as proportional to the difference between the reinforcement supported by the actual outcome and the total expectation generated by cues present on that trial. A standard notation is ΔV_i = α_i β(λ − ΣV_present), where V_i is associative strength for cue i, α_i indexes its learning effectiveness, β denotes a reinforcement-related rate parameter, λ represents asymptotic support from the actual outcome, and ΣV_present is the combined expectation. The expression is a model of trial-wise associative updating; its parameters are theoretical variables requiring careful operational estimation Rescorla & Wagner (1972).
Blocking follows directly: after A predicts the outcome well, the compound AX generates approximately the expected result, leaving little error for X. Overshadowing follows when elements compete in acquiring strength across compound trials. When an expected outcome is withheld, negative error reduces excitatory prediction and can support inhibitory learning depending on the training history. The model consequently connects different conditioning phenomena with a common learning rule, explaining its extraordinary importance for later connectionist and computational accounts Rescorla & Wagner (1972) Miller et al. (1995).
Important limitations concern the representation of stimuli, elapsed time, changing attention, and learning without direct outcome delivery. The simplest model treats events as trial units, provides limited representation of temporal microstructure, and has difficulty with discriminations requiring configural or sequential structure. Miller, Barnet, and Grahame's appraisal demonstrates the difference between explanatory influence and comprehensive empirical adequacy. The question is not whether the learning rule is historically important—it is—but which observations require modifications or competing mechanisms Miller et al. (1995) Pearce (1994) Harris (2025).
Conditioned Inhibition and Learning from an Omitted Outcome
Conditioned inhibition concerns learning that a stimulus signals a reduction in the likelihood of an otherwise expected event. A widely used feature-negative discrimination presents A followed by reinforcement, and AX without reinforcement. Within that arrangement X can come to predict omission in the presence of A. Its inhibitory status should be tested rather than inferred solely from weak responding: summation testing asks whether X reduces the response controlled by another independently trained excitor, while retardation testing asks whether X is slow to become excitatory under later reinforcement Rescorla (1969) Rescorla & Wagner (1972).
Inhibitory control exposes the contextual character of expectation. A cue carrying negative predictive information usually does so relative to an excitatory background, a target stimulus, or a temporal window. Consequently a response deficit during X alone may be difficult to interpret when the participant never expects the outcome in that condition. A systematic design distinguishes a genuine inhibitory relation from mere nonreinforcement, latent inhibition, habituation, and a stimulus whose sensory properties suppress the measurement process Rescorla (1969) Bouton (2004).
Theoretical models make different claims about inhibitory representation. Rescorla–Wagner formalism treats negative associative value as a counterpart of positive prediction; attentional models ask whether unexpected omissions increase processing of relevant cues; contextual retrieval theories treat the learned omission as conditional on circumstances. These explanations converge on the empirical requirement that negative evidence can carry new information about the environment. Such evidence also clarifies why avoidance of expected threat is an acquisition problem in its own right and why extinction must be analyzed separately from simple erasure Pearce & Hall (1980) Bouton (2004).
Prediction Error and the Experimental Logic of Surprise
An outcome is surprising when it differs from the prediction supported by the organism's current model of the situation. Surprise can concern whether the event occurs, when it occurs, its magnitude, or its sensory identity. Conditioning experiments treat surprise as a proposed process influencing learning rather than a synonym for experimenter-defined novelty. The same physical outcome may generate different learning if past training has altered what the organism anticipates Rescorla & Wagner (1972) Niv & Schoenbaum (2008).
Surprise-based explanations gain traction when the experimenter can independently alter prior prediction and observe consequences for acquisition. Blocking and unblocking arrangements provide such tests by comparing compound trials that preserve or change the expected outcome. Experimental conclusions are strongest when the newly informative cue is tested separately and when changes in attention, motivation, and processing demands are controlled. Neural recordings can offer converging evidence, but similar neural time courses across tasks do not identify a unique cognitive model without explicit model comparison Schultz et al. (1997) Niv & Schoenbaum (2008).
Contemporary computational research separates signed errors, which indicate a discrepancy with direction, from unsigned surprise, which can regulate learning rate or attention. A reward omission and an unexpectedly large reward may both change processing, yet they can require different associations and behaviors. The distinction is important for integrating attention theory with error correction and for preventing the term prediction error from becoming an uninformative label applied to any experimental deviation Pearce & Hall (1980) Gershman (2015).
Mackintosh's Attentional Theory and Predictive Validity
Mackintosh's 1975 account proposed that learning effectiveness is itself altered by experience. A cue that predicts reinforcement more accurately than concurrent cues may gain associability, while redundant or unreliable cues may lose it. This process allows selective attention to be expressed within a quantitative framework rather than assuming that all physically present stimuli receive identical processing. The theory links discrimination learning, cue competition, and the consequences of prior relevance training, with important implications for the generalization of acquired control Mackintosh (1975).
The attentional interpretation of blocking emphasizes that once A reliably predicts reinforcement, another element may be treated as redundant and processed less effectively. Its predictions can be examined by asking whether earlier experience with predictive and nonpredictive cues changes subsequent learning about those cues in a different phase. Appropriate testing separates initial cue salience from learning-induced changes in attention, because the two can have similar effects on observed acquisition. In animal preparations this requires counterbalanced stimulus identities and explicit tests of learned relevance Mackintosh (1975) Wagner et al. (1968).
Selective attention is not a complete theory of what a signal represents. A learner could attend strongly to a cue yet still assign it incorrect outcome identity or temporal location. Conversely, the cue could carry a relational representation that is expressed weakly because other cues dominate behavioral control. The best contemporary use of the Mackintosh tradition is to generate falsifiable predictions about processing history and learning rates while allowing distinct explanations of outcome content and performance Pearce & Hall (1980) Harris (2025).
Pearce–Hall Associability and Uncertainty
Pearce and Hall's 1980 theory offers a different principle for associability: stimuli that have been followed by uncertain or poorly predicted outcomes retain high learning effectiveness, whereas accurately predicted outcomes reduce the stimulus's capacity to support additional learning. The theoretical focus is therefore on uncertainty about consequences rather than preferential attention to whichever signal is already the best predictor. This creates an empirically consequential contrast with the Mackintosh approach Pearce & Hall (1980).
The distinction becomes sharp in experimental transitions. An established reliable signal may command attention under one theory because it is predictively superior, yet lose associability under another because its consequence is already expected. When an outcome changes unexpectedly, Pearce–Hall reasoning predicts renewed processing and greater learning in subsequent trials. Experimental adjudication requires procedures distinguishing attentional orienting from subsequent associative change, not merely the use of the term uncertainty in retrospect Pearce & Hall (1980) Mackintosh (1975).
The models can also be read as capturing different aspects or phases of adaptive learning, and modern formal approaches allow a learner to adjust its learning rate as uncertainty changes. However, unifying theories need independent evidence for which computational variable corresponds to which behavioral or neural signal. A general tendency for an uncertain cue to attract attention does not by itself establish a particular equation, while reliable asymmetries between prediction accuracy and surprise help constrain the model space Gershman (2015) Harris (2025).
Temporal Order, Interval Learning, and Predictive Time
A cue's usefulness depends on when the outcome is expected. Delay conditioning presents the cue before the US with temporal overlap; trace conditioning inserts an interval between cue offset and US; other procedures place the US before the cue or distribute events across longer temporal gaps. These arrangements alter the predictive value and the memory demands of the cue. Acquisition should therefore be studied in relation to the interval, its variability, the ongoing background rate, and the temporal form of conditioned responding Pavlov (1927) Gallistel & Gibbon (2000).
Timing models treat conditioning as learning a temporal distribution or map rather than acquiring only a scalar associative strength. Gallistel and Gibbon's analysis emphasized event intervals and rates, and Balsam and Gallistel examined evidence for temporal maps and informativeness. Such approaches clarify why trial number alone is an inadequate measure of experience and why the information gained from a cue must be evaluated relative to what could be expected in its absence Gallistel & Gibbon (2000) Balsam & Gallistel (2009).
An experiment can separate these accounts by holding cue–outcome pairings constant while changing cue duration, intertrial spacing, background outcome rate, or variability in time-to-outcome. Time-specific anticipatory responding provides a richer dependent measure than total response magnitude. Contemporary modeling recognizes that discrete-time error-correction systems can also represent time when provided suitable states; consequently temporal evidence discriminates among specified implementations, rather than categorically excluding all prediction-error models Harris (2025) Qian et al. (2025).
Rates, Background Context, and Information Theory
The background context is an active part of the conditioning problem because outcomes can occur outside the target CS. When their unsignaled rate rises, the presence of a paired cue may add progressively less information about what will happen next. The information-theoretic perspective formalizes this dependence on uncertainty reduction and expected event timing. It treats temporal spacing and contingency as components of an environmental prediction problem in which relevant baselines must be measured rather than assumed Balsam & Gallistel (2009) Harris (2025).
Context is sometimes operationalized as a static set of physical features and sometimes as a state carrying a history of elapsed time, recent outcomes, and latent task conditions. These alternatives yield different computational predictions. A learner could respond less to a cue when unsignaled outcomes become common because the contextual expectation has increased, because the inferred causal status of the cue has declined, or because a state representation now allocates prediction across different moments. Experimental interventions that mark additional outcomes with another independent cue help discriminate between explanations Qian et al. (2025).
Information-theoretic models gain precision through explicit specification of event rates, intervals, and uncertainty measures. They also face an explanatory requirement: information is an analyst-defined quantity, while behavior is generated by particular memory and decision processes. Showing that informational value tracks the learning curve is evidence of functional alignment; demonstrating the mechanism requires experiments that vary putative representations and processing parameters separately. The result is an integrated but differentiated research agenda for prediction, timing, and contingency Gallistel & Gibbon (2000) Gershman (2015).
Configural Learning and Conditional Discriminations
Some discriminations require animals to respond to combinations rather than isolated components. Positive-patterning training can reinforce a compound while leaving its elements nonreinforced, whereas negative-patterning training reverses the relation. An exclusively additive interpretation of elemental associative strengths has difficulty with particular patterning arrangements, motivating configural and connectionist accounts. Such tests reveal what counts as a stimulus representation rather than only how rapidly its value changes Pearce (1994).
Configural theories propose that a compound is represented as an organized whole or as a collection of unique features, and that responding to individual elements depends on similarity to learned representations. This can account for substantial variation in generalization and discrimination gradients. Elemental theories can sometimes be extended with unique cues, inhibitory features, or flexible attention; comparing predictions therefore requires precise model implementations rather than broad verbal labels. The strongest data include transfer to novel combinations whose similarity relations have been independently manipulated Pearce (1994) Miller et al. (1995).
The implications extend to practical prediction: complex situations can become meaningful signals whose consequences are determined by arrangements of cues, contextual relationships, and previously learned discriminations. Learning about one visual element in a device, another contextual element in an environment, and their combination may not be separable. Yet importing the construct into human instruction, interface training, or social interpretation requires empirical tests of whether human participants treat those features configurally or propositionally in the target domain Pearce (1994) Lovibond & Shanks (2002).
Outcome Identity, Revaluation, and What a Cue Predicts
Conditioned responding can reveal information about the particular outcome expected rather than only its general desirability. In an outcome-specific design, different cues predict different foods or other consequences; subsequent changes to the outcome can alter how the cue controls behavior. Such tests motivate representations that contain sensory properties, motivational attributes, and timing, since a scalar measure of associative strength alone does not preserve which event is anticipated. The observation that a cue has acquired value must therefore be separated from the stronger hypothesis that it activates a representation of an outcome's identity Rescorla (1988) Kahnt & Schoenbaum (2025).
Revaluation procedures are theoretically powerful because they change what the outcome means without necessarily presenting the cue again. If behavior elicited by the cue changes at test, one possibility is that it retrieves a representation whose current value can be recomputed. However, outcome revaluation is also affected by performance rules, motivational state, and direct exposure to the altered outcome. Reliable interpretation requires testing the cue alone, comparing different outcomes, and controlling the degree of habituation or new learning produced by the revaluation manipulation itself Rescorla (1988) Niv & Schoenbaum (2008).
Current neural scholarship reinforces the significance of specific event content. Kahnt and Schoenbaum's 2025 synthesis argues that some dopaminergic transients reflect information about reward features beyond scalar utility, complicating the claim that all relevant dopamine teaching signals encode one-dimensional value. The scientific task is to specify whether a task manipulates value, identity, category, timing, or inferred causal structure, then ask which computational predictions follow. This distinction links the behavioral science of conditioning with theories of semantic representation and goal-directed planning while retaining its own experimental methods Kahnt & Schoenbaum (2025).
Human Contingency Judgment and the Role of Awareness
Human participants can report their estimates of whether a signal predicts an outcome, and these reports offer a second observable beside physiological or motor conditioning. The strongest experiments compare judged contingencies with autonomic reactions, conditioned behavior, and direct knowledge tests. A participant who reports the cue–outcome relation accurately may still express weak physiological discrimination; conversely, a physiological difference does not automatically establish the absence of propositional knowledge. Lovibond and Shanks surveyed evidence that awareness matters substantially across human conditioning paradigms, while emphasizing problems in how awareness is measured Lovibond & Shanks (2002).
Methodological disputes often arise from the timing and sensitivity of the awareness measure. Retrospective verbal questions may miss knowledge available earlier in training; concurrent reports can change attention and strategy; dichotomous awareness classifications can discard partial or uncertain knowledge. The dependent measures also have different reliability and time scales, and an apparently unconscious effect can be sensitive to choice of statistical criterion. Claims of awareness-independent conditioning require explicit standards for equivalence or absence of relevant knowledge and a task design that makes reporting feasible Lovibond & Shanks (2002) Lonsdorf et al. (2017).
Lovibond and Westbrook's 2025 review places expectancy at the center of both acquisition and conditioned performance, noting that instruction about an impending US may elicit anticipatory reactions even when a person has not learned through direct pairings. Their synthesis highlights the need to model how propositional information, experience, and bodily preparation interact. Instructional acquisition also provides an experimental tool: when the learned prediction is altered by a verbal statement, subsequent behavior can be contrasted with the effect of directly experienced prediction error Lovibond & Westbrook (2025).
Causal Inference, Interventions, and Learned Relations
Learning that X precedes Y can support prediction without establishing that X causes Y. Causal reasoning requires distinguishing observations of events from interventions on events, which may yield different conclusions even when the observed contingency is similar. Blaisdell and colleagues tested this distinction in rats using a preparation designed to compare interventions with observations. Their findings support the relevance of causal-model hypotheses for some animal behavior while leaving open the representational complexity and generality of the mechanism Blaisdell et al. (2006).
Beckers and colleagues investigated forward blocking in rats under conditions sensitive to causal-inference constraints. Such tests matter because a decrease in learning about a redundant cue can be explained by different internal operations: reduced associability, shared outcome prediction, or an inferred causal structure in which an added event is irrelevant. An experiment can favor a causal interpretation when controlled interventions alter expectations as the proposed model predicts and alternative elemental accounts have been fitted and tested explicitly Beckers et al. (2006) Rescorla & Wagner (1972).
Human causal judgment also carries assumptions about how many causes are sufficient, whether causes compete or combine, and whether causes are merely diagnostic symptoms of another process. The resulting judgments may differ across observational and intervention conditions and with changes in prior knowledge. The important methodological principle is to report the causal graph or generative assumptions being tested and to compare the model against a well-specified associative alternative. Causal language in the investigator's prose cannot substitute for experimental evidence that the participant represented causal intervention Gershman (2015) Lovibond & Westbrook (2025).
Bayesian Learning, Latent Causes, and Uncertainty Updating
Probabilistic models describe how a learner updates beliefs about cue–outcome relations as evidence accumulates. Bayesian accounts represent uncertainty explicitly and can distinguish the weight assigned to one new observation from the current strength of the inferred relation. Such theories offer an account of why surprising events have different effects depending on how variable the environment is believed to be. Gershman's probabilistic synthesis illustrates how several classical learning phenomena can be reconstructed through inference over causes and latent environmental states Gershman (2015).
A latent-cause model proposes that the learner groups experiences into inferred situations, each of which can support different predictions. An outcome omission following a cue may indicate a genuine change in the world, a new temporary context, or noise within an existing state. These alternatives have different implications for later recovery and renewed responding. The model must specify how observations are assigned to latent states and how prior expectations are updated; otherwise it can accommodate any result through post hoc state assignment Gershman (2015) Bouton (2004).
Statistical fitting alone provides limited causal explanation when models have many flexible parameters or unobserved states. Cross-validation, parameter recovery, model recovery, out-of-sample behavioral prediction, and targeted interventions can separate genuine explanatory improvements from fit to noise. A scientifically useful comparison also measures the stimulus timing and background event distribution actually experienced by participants. This is particularly important when Bayesian models are compared with associative algorithms that have been granted less elaborate state representations Harris (2025) Qian et al. (2025).
Expectation, Response Generation, and Motivational State
A predictive representation does not uniquely determine the bodily response that follows. The same cue–outcome expectancy may support food-cup approach, cue-directed interaction, orienting, swallowing preparation, or inhibitory restraint as a function of outcome imminence and the organism's response repertoire. Theoretical attention to expectation therefore shifts part of the explanatory burden from acquisition of information to selection and organization of anticipatory behavior. Rescorla's relational treatment and Lovibond and Westbrook's recent synthesis converge in making that distinction central Rescorla (1988) Lovibond & Westbrook (2025).
The distinction helps interpret situations where a cue is clearly informative but generates little measured CR. Motivational satiation, action competition, ceiling or floor effects, or weak measurement sensitivity may conceal preserved knowledge. Conversely, a change in performance could be generated by altered motivation without new learning about contingency. Outcome-specific and cross-response tests, direct probability judgments where possible, and tests under changed motivational conditions can clarify the degree to which acquisition and expression have been separated Bouton & Moody (2004) Lonsdorf et al. (2017).
The problem also has implications for studies of individual difference and human–artificial interaction. A person may understand a warning yet behave inconsistently because competing goals, time pressure, learned habits, or social factors shape response production. Such transfer from laboratory conditioning to applied settings requires direct empirical support; it cannot be inferred from the existence of a predictive-learning equation. The fundamental psychological principle is that learned environmental expectation interacts with response-selection systems whose dynamics must be studied independently McNally et al. (2011) Lovibond & Westbrook (2025).
Dopamine, Temporal-Difference Learning, and Prediction Signals
Reward-related midbrain dopamine responses supplied one of the most influential bridges between behavioral conditioning and formal learning theory. Schultz, Dayan, and Montague described neural activity consistent with differences between expected and obtained reward, including changes when a previously unexpected reward became predicted by a cue. Temporal-difference models formalize error signals across successive predictive states, enabling learning before the eventual outcome occurs. The empirical contribution is a measurable physiological signal with a functional relationship to reward prediction; its causal and representational interpretation requires additional experiments Schultz et al. (1997) Niv & Schoenbaum (2008).
Neural signals are shaped by a richer range of variables than a simple reward-minus-value equation might suggest. Task beliefs, movement, sensory identity, reward timing, novelty, motivational state, and anatomical recording location all matter to the measurement. Kahnt and Schoenbaum's review of reward-feature learning cautions against reducing dopaminergic teaching to a single value dimension in every paradigm. In a scientifically disciplined account, a neural prediction-error signal is a proposed computational variable with specific tests, and the measured fluorescence or spiking response is an observation whose relationship to that variable must be estimated Kahnt & Schoenbaum (2025).
From the perspective of behavioral learning, neural evidence becomes most discriminative when the experiment manipulates predictions while matching reward delivery as closely as possible. It also matters whether a putative prediction error precedes, accompanies, or follows the change in behavior. Correlated activity may indicate an input to learning, a readout of an inferred state, or an effect of another process. Converging perturbations and computational predictions provide stronger evidence about a circuit's causal role than a characteristic average response alone McNally et al. (2011) Qian et al. (2025).
Prospective Contingency in the 2025 Mouse Dopamine Experiments
Qian and colleagues performed a Pavlovian contingency-degradation investigation in mice while measuring dopamine-related responses in ventral striatal axons. Their crucial comparison added rewards without a preceding cue in one condition and introduced cues for those extra rewards in another. Anticipatory licking and cue-related dopamine activity were reduced when extra rewards were uncued, whereas both remained relatively stable when the added rewards were signaled. The evidence extends beyond the familiar assertion that more unsignaled rewards impair conditioning: it shows that the predictive organization of the extra outcomes matters even when their physical occurrence is held in view Qian et al. (2025).
The authors evaluated traditional contingency formulations, a newer associative causal model, and temporal-difference models with alternative representations of the intertrial interval. Their analyses found that temporal-difference algorithms with appropriately constructed belief-state representations accounted for the behavioral and dopamine patterns, while simple state descriptions did less well. Recurrent neural networks trained within a temporal-difference framework developed state organizations resembling their effective hand-specified model. This result illustrates how a formal prediction-error learning rule can appear inadequate under one state encoding and predictive under a more appropriate description of the organism's prospective information Qian et al. (2025).
The study is a strong case of combined intervention, neural measurement, and computational adjudication, although it does not establish a universal mechanism for all conditioning forms, species, or motivational systems. Model performance was assessed within a particular appetitive mouse preparation, with specific neural recordings and event schedules. The broader methodological lesson is to specify which states a learner can infer and what predictions those states generate before declaring two learning theories empirically incompatible. Behavioral contingency and dopamine signals here constrain models jointly, not interchangeably Qian et al. (2025) Harris (2025).
Defensive Conditioning and the Organization of Threat
In aversive conditioning, a previously neutral cue comes to predict a harmful or unpleasant event. Investigators measure defensive immobility, startle potentiation, autonomic changes, avoidance-related preparation, and subjective expectancy depending on the preparation. These outcomes are governed by overlapping but distinguishable systems, so the label fear conditioning should not be treated as a direct measure of one subjective state. McNally, Johansen, and Blair integrated prediction-based learning with work on the neural circuitry of conditioned defensive responding McNally et al. (2011).
Human laboratory threat paradigms introduce specific measurement challenges. Lonsdorf and colleagues discuss stimulus selection, US calibration, reinforcement schedules, verbal instruction, habituation, physiological nonresponders, trialwise analyses, and interpretation of extinction and return-of-fear tests. Skin conductance, startle, ratings, and neural measurements should be reported as operationally distinct endpoints. Differences among them are scientifically informative, because cognitive knowledge of danger and biological defensive expression can be dissociated under some conditions Lonsdorf et al. (2017).
Clinical extrapolation requires additional caution. Laboratory acquisition may model selected components of threat prediction, yet a specific anxiety disorder involves development, learning history, appraisal, regulation, social context, and functional impairment across time. Conditioning experiments provide causal probes of some mechanisms; they are not diagnostic tests for an individual. Work on extinction and relapse informs conceptual models of exposure-based interventions, while treatment recommendations require disorder-specific clinical trials and current professional guidelines beyond this foundational dossier Vervliet et al. (2013) Dunsmoor et al. (2015).
Extinction, New Learning, and the Survival of Earlier Predictions
Extinction commonly refers to a decline in conditioned responding when a previously predictive cue is presented without its expected outcome. Its theoretical importance is that reduced performance can arise from new learning about omission while an earlier acquired relation remains available. Recovery phenomena after passage of time, context change, or unsignaled outcome presentations demonstrate that one endpoint immediately after extinction cannot characterize what was retained. Bouton's work explains how context and retrieval conditions can govern which learning controls behavior Bouton (2004).
Modern analyses of extinction integrate error correction, inhibitory relations, temporal structure, and latent-state inference. Dunsmoor and colleagues argue for attention to the different ways an organism updates threat predictions during extinction and how those updates are expressed. These accounts do not imply that every return of responding has one cause; rather, they motivate experimentally differentiating renewal, spontaneous recovery, reinstatement, and reacquisition. A full treatment of those mechanisms belongs in the neighboring extinction dossier; their role here is to show what cue–outcome learning represents Dunsmoor et al. (2015) Vervliet et al. (2013).
Extinction is also a stringent test for the distinction between associative value and explicit expectation. A participant may correctly anticipate that an outcome will no longer occur while a conditioned physiological response persists temporarily, and laboratory measurements can differ in their rate of change. Investigators should specify response endpoints, contexts, intervening events, and delayed follow-up conditions to decide whether an observed decline reflects new information, changed motivation, altered retrieval, or statistical instability. This resolves the central claim at the level of testable evidence rather than a metaphor of forgotten fear Lonsdorf et al. (2017) Lovibond & Westbrook (2025).
Neural Circuits, Measurement Levels, and Causal Claims
The conditioned prediction may be described at several connected levels: the environmental schedule, a behavioral estimate of expected outcome, a formal computational variable, activity within neural populations, and the peripheral or overt response. These levels answer different questions. A contingency manipulation establishes a change to the input statistics; a conditioned response shows altered behavior; a circuit perturbation can test a neural contribution; a model specifies a proposed algorithm. Moving from one level to another requires evidence connecting the measures rather than merely shared terminology McNally et al. (2011) Qian et al. (2025).
Threat conditioning often motivates study of amygdala-centered circuits, while reward learning frequently emphasizes dopamine projections and striatal systems. Other preparations depend on cerebellar timing and species-specific motor circuitry, but the present dossier focuses on the psychological learning relations that such systems instantiate. The same formal concept may have different physiological implementations, and one circuit can contribute to several aspects of learning, attention, response expression, or valuation. Neural specialization should therefore be established by experimentally separating functions, not by assigning the entire psychological construct to one region Schultz et al. (1997) McNally et al. (2011).
Combining physiology and behavior improves inference when both are measured at appropriate resolution and analyzed with prespecified competing predictions. A neural signal may track an outcome's expected occurrence, uncertainty, identity, or the participant's belief about a hidden state; these hypotheses require different interventions and model comparisons. The rich 2025 data on contingency-sensitive dopamine responses exemplify this standard. Research on neural implementation continues to develop in the neighboring dossier Neural Mechanisms of Learning and Prediction, while the canonical subject here remains experimentally established cue–outcome learning Qian et al. (2025) Kahnt & Schoenbaum (2025).
Comparative Methods and Species-Appropriate Designs
Pavlovian conditioning has been observed through different behavioral repertoires, but comparable experimental labels do not guarantee equivalent psychological measurements across species. Salivation in dogs, approach to food-associated cues in rodents, autonomic responses in humans, and changes in eye movement or timing can all reflect acquired predictions. To compare them, researchers must establish the species-specific salience of the signal, the biological importance of the outcome, the relevant motor opportunities, and the appropriate baseline of behavior. The design should preserve the causal structure of the experimental manipulation while acknowledging distinct response channels Pavlov (1927) Rescorla (1988).
Cross-species inference requires more than observing a shared acquisition curve. Distinct sensory ranges and temporal horizons may change the representation of a compound cue; ecological experience may alter prior expectations; and reinforcement histories can affect responsiveness. A proposed conserved prediction-error computation becomes more persuasive if species-appropriate tasks reproduce a diagnostic pattern under interventions that distinguish it from a simpler salience or response-expression account. Conversely, species differences are valuable findings when methodological equivalence has been demonstrated Miller et al. (1995) Harris (2025).
Animal experiments also raise ethical questions about the intensity and justification of aversive outcomes. Study designs should minimize unnecessary exposure, select scientifically justified and species-appropriate reinforcers, and follow the applicable institutional ethics and animal welfare rules. Nonaversive reward-learning tasks can answer many theoretical questions, and human research should ensure consent, transparent risk disclosure within justified deception protocols, and suitable debriefing. Scientific rigor includes ethical design because uncontrolled distress and motivation can themselves compromise interpretation Lonsdorf et al. (2017) Qian et al. (2025).
Validity, Statistical Inference, and Reproducibility
Learning curves are repeated observations nested within individuals, commonly with substantial heterogeneity in starting response, acquisition rate, and asymptote. Analyses that rely only on a final group average can conceal informative shifts in timing or strategy and may inflate apparent certainty if observations are treated as independent. Appropriate statistical approaches can model participant-level differences, trial dependencies, and uncertainty intervals, while transparently reporting exclusions, missing data, and the handling of extreme response values. The statistical model should correspond to the experimental unit and to the prediction of the theory being evaluated Lonsdorf et al. (2017) Harris (2025).
Construct validity requires demonstrating that the dependent measure is sensitive to the intended process and that alternative sources of the effect have been addressed. A blocking result should include adequate evidence that A had acquired predictive control, that X was perceptually available, that comparison stimuli had appropriate training, and that test conditions did not change motivation or response competition asymmetrically. A conditioned-inhibition claim needs explicit inhibitory tests. A neural model needs a link between its latent variables and recorded data. These demands make replication a test of explanatory stability across scientifically relevant variations Rescorla (1969) Miller et al. (1995) Qian et al. (2025).
Evidence synthesis should preserve the boundary between demonstrable replication and interpretive consensus. Foundational experiments are historically decisive in motivating models, but their original measurement and sample choices can differ greatly from current standards. Recent reviews sharpen model contrasts yet remain reviews unless their procedures justify a systematic-review designation. This manuscript uses a critical narrative method, privileges original task descriptions where accessible, records disagreements without averaging incompatible designs, and treats the absence of a universal theory as a continuing scientific problem Harris (2025) Lovibond & Westbrook (2025).
Model Comparison and the Problem of Psychological Identifiability
Several models can generate similar learning curves through very different assumptions. A trial-wise associative update, an attentional adjustment, a temporal-rate comparison, and a Bayesian latent-state inference may all predict faster responding when a cue is highly reliable. The task of model comparison is to find experimental conditions where their predictions diverge, such as cue competition after prior predictive training, unexpectedly altered outcome identity, changes in background reward rate, or novel temporal structures. Comparisons must be made on common datasets and with comparable information supplied to each model Rescorla & Wagner (1972) Gallistel & Gibbon (2000) Gershman (2015).
Parameter identifiability is a separate problem. Even within one learning algorithm, multiple parameter settings can produce nearly identical observations, especially when acquisition is measured only at the group level or a narrow range of trials. Recovery exercises use simulations to ask whether the available design can distinguish proposed parameter values or separate competing algorithms at all. Held-out predictions help evaluate generalization, but they do not rescue a model whose hidden states and observables were defined after seeing the result. Strong evidence requires prospective predictions and explicit design choices Harris (2025) Qian et al. (2025).
Model complexity is justified when it captures theoretically necessary distinctions with successful predictive consequences. Qian and colleagues illustrate this by showing that state representation changes whether a temporal-difference learner can account for the difference between cued and uncued extra rewards. The addition of latent state is scientifically interpretable because it corresponds to the prospective information available in the task and predicts behavior and dopamine measures jointly. This illustrates a general norm: compare the theories as fully specified candidate processes, with independent evidence constraining their representations Qian et al. (2025).
Artificial Learners, Predictive Systems, and the Artificial Era
Computational conditioning theories have long shaped machine-learning concepts, particularly temporal-difference updating and attention to diagnostic features. In the Artificial Era, this cross-disciplinary relation supports controlled comparison between biological contingency learning and artificial systems that infer event probabilities, update predictive states, and plan on the basis of environmental feedback. Such comparisons are methodological: similar learning curves can arise from different architectures, training regimes, objectives, and memory structures. A valid study therefore specifies the artificial system's input representation, training data, action interface, evaluation set, and update rule, then compares concrete behavioral predictions Rescorla & Wagner (1972) Qian et al. (2025).
A particularly informative cross-system test would replicate a contingency-degradation design with both uncued extra outcomes and extra outcomes predicted by their own signals. One could then measure whether the artificial learner adjusts the target cue's predictive value when the overall outcome rate changes, whether it distinguishes causal intervention from observation, and how it handles temporal and compound cues. The comparison could reveal algorithmic analogies with laboratory learning while permitting direct inspection of artificial state and parameter updates. Proposed architectures should be evaluated on unseen conditions rather than recognized only through superficial agreement with familiar effects Qian et al. (2025) Blaisdell et al. (2006).
The philosophical horizon From Homo to Artificial makes questions of independent nonbiological learning and world representation historically salient, but an observed computational similarity does not itself resolve questions of consciousness, sentience, agency, or personal identity. These are distinct theoretical and empirical research problems requiring their own operational criteria. Within Psychology by Bogdanova, the purpose of the comparison is to specify transferable research methods and genuine behavioral mechanisms while preserving the evidentiary distinction between biological observation, engineered implementation, and philosophical interpretation Kahnt & Schoenbaum (2025) Gershman (2015).
Integrated Findings and Unresolved Research Problems
The best-supported general conclusion is that Pavlovian conditioning depends on the predictive relation among cues, outcomes, time, and background conditions, with expression mediated by the organism's physiological and behavioral systems. Blocking and overshadowing reveal cue competition; conditioned inhibition establishes the importance of expected omissions; temporal manipulations show the relevance of when an event is predicted; attentional studies demonstrate changes in cue processing; and human studies identify the importance of expectancy and instruction. Each finding constrains the space of theories without uniquely selecting a universal mechanism Rescorla (1968) Miller et al. (1995) Harris (2025) Lovibond & Westbrook (2025).
Five research challenges remain especially consequential. Scientists must distinguish learning from expression with better process measures; establish how temporal state and contextual uncertainty are represented; adjudicate between attentional, elemental, configural, and inferential explanations through interventions; link neural teaching signals to learning causally and with content specificity; and determine which mechanisms generalize across species, individuals, and naturalistic settings. Carefully designed preregistered comparisons, robust participant-level analyses, independent replications, and open computational implementations can narrow these uncertainties Lonsdorf et al. (2017) Qian et al. (2025) Kahnt & Schoenbaum (2025).
Progress will likely come from experimentally linking distinct observational levels rather than insisting that one term—association, prediction error, attention, or inference—captures the whole process. A program that measures behavior, expectancy, event timing, and physiological signals under controlled changes in contingency can test whether a common computation explains more than one readout. This approach also permits principled comparison with engineered artificial learners, treating cross-system generality as a hypothesis rather than an assumption. The resulting scientific picture is a plural but connected account of how events acquire anticipatory meaning Gershman (2015) Harris (2025).
Conclusion
Pavlovian conditioning is a fundamental mechanism of adaptive prediction in which signals acquire behavioral significance through their relations to consequential events. The history from Pavlov's physiological preparations to formal learning theory and contemporary neural experimentation shows an expanding account of what the learner acquires: predictive outcome probability, timing, cue competition, omission, causal and configurational structure, and state-dependent expectation. The contemporary object of study is a system of learned relations whose expression depends on context, motivation, and the availability of suitable responses Pavlov (1927) Rescorla (1988) Lovibond & Westbrook (2025).
The empirical foundation supports a central analytical distinction. The programmed contingency is a property of an experiment; the acquired representation is a hypothesis about the learner; the observed conditioned response is the evidence produced by the whole behaving system. Prediction-error, attentional, temporal, and probabilistic theories offer complementary and competing ways to explain transitions among these levels. Their value rests on precision, discriminating experiments, and independent replication. Current research on cue-specific dopamine dynamics further demonstrates the power of jointly testing behavioral and neural predictions while specifying the learner's inferred state Harris (2025) Qian et al. (2025).
The scientific significance extends from comparative learning and human expectancy to defensive regulation, cognitive theories of causal inference, and experimentally controlled studies of artificial learning systems. These extensions remain accountable to the same methodological principle: specify the environment, identify the proposed representation, measure behavior and relevant physiology, and test alternative explanations. Pavlovian conditioning and contingency learning accordingly remain a foundational research program for understanding how agents acquire, revise, and use predictions about the world Rescorla (1988) Gershman (2015).
References
- Balsam, P. D., & Gallistel, C. R. (2009). Temporal maps and informativeness in associative learning. Trends in Neurosciences, 32(2), 73–78. https://doi.org/10.1016/j.tins.2008.10.004
- Beckers, T., Miller, R. R., De Houwer, J., & Urushihara, K. (2006). Reasoning rats: Forward blocking in Pavlovian animal conditioning is sensitive to constraints of causal inference. Journal of Experimental Psychology: General, 135(1), 92–102. https://doi.org/10.1037/0096-3445.135.1.92
- Blaisdell, A. P., Sawa, K., Leising, K. J., & Waldmann, M. R. (2006). Causal reasoning in rats. Science, 311(5763), 1020–1022. https://doi.org/10.1126/science.1121872
- Bouton, M. E. (2004). Context and behavioral processes in extinction. Learning & Memory, 11(5), 485–494. https://doi.org/10.1101/lm.78804
- Bouton, M. E., & Moody, E. W. (2004). Memory processes in classical conditioning. Neuroscience & Biobehavioral Reviews, 28(7), 663–674. https://doi.org/10.1016/j.neubiorev.2004.09.001
- Dunsmoor, J. E., Niv, Y., Daw, N., & Phelps, E. A. (2015). Rethinking extinction. Neuron, 88(1), 47–63. https://doi.org/10.1016/j.neuron.2015.09.028
- Gallistel, C. R., & Gibbon, J. (2000). Time, rate, and conditioning. Psychological Review, 107(2), 289–344. https://doi.org/10.1037/0033-295X.107.2.289
- Gershman, S. J. (2015). A unifying probabilistic view of associative learning. PLOS Computational Biology, 11(11), e1004567. https://doi.org/10.1371/journal.pcbi.1004567
- Harris, J. A. (2025). Modelling the acquisition of Pavlovian conditioning. Neurobiology of Learning and Memory, 219, 108059. https://doi.org/10.1016/j.nlm.2025.108059
- Kahnt, T., & Schoenbaum, G. (2025). The curious case of dopaminergic prediction errors and learning associative information beyond value. Nature Reviews Neuroscience, 26, 169–178. https://doi.org/10.1038/s41583-024-00898-8
- Lonsdorf, T. B., Menz, M. M., Andreatta, M., Fullana, M. A., Golkar, A., Haaker, J., Heitland, I., Hermann, A., Kuhn, M., Kruse, O., Meir Drexler, S., Meulders, A., Nees, F., Pittig, A., Richter, J., Römer, S., Shiban, Y., Schmitz, A., Straube, B., ... Merz, C. J. (2017). Don't fear 'fear conditioning': Methodological considerations for the design and analysis of studies on human fear acquisition, extinction, and return of fear. Neuroscience & Biobehavioral Reviews, 77, 247–285. https://doi.org/10.1016/j.neubiorev.2017.02.026
- Lovibond, P. F., & Shanks, D. R. (2002). The role of awareness in Pavlovian conditioning: Empirical evidence and theoretical implications. Journal of Experimental Psychology: Animal Behavior Processes, 28(1), 3–26. https://doi.org/10.1037/0097-7403.28.1.3
- Lovibond, P. F., & Westbrook, R. F. (2025). The role of expectancy in Pavlovian conditioning. Psychological Review, 132(6), 1275–1293. https://doi.org/10.1037/rev0000516
- Mackintosh, N. J. (1975). A theory of attention: Variations in the associability of stimuli with reinforcement. Psychological Review, 82(4), 276–298. https://doi.org/10.1037/h0076778
- McNally, G. P., Johansen, J. P., & Blair, H. T. (2011). Placing prediction into the fear circuit. Trends in Neurosciences, 34(6), 283–292. https://doi.org/10.1016/j.tins.2011.03.005
- Miller, R. R., Barnet, R. C., & Grahame, N. J. (1995). Assessment of the Rescorla-Wagner model. Psychological Bulletin, 117(3), 363–386. https://doi.org/10.1037/0033-2909.117.3.363
- Niv, Y., & Schoenbaum, G. (2008). Dialogues on prediction errors. Trends in Cognitive Sciences, 12(7), 265–272. https://doi.org/10.1016/j.tics.2008.03.006
- Pavlov, I. P. (1927). Conditioned reflexes: An investigation of the physiological activity of the cerebral cortex (G. V. Anrep, Trans.). Oxford University Press. https://psychclassics.yorku.ca/Pavlov/
- Pearce, J. M. (1994). Similarity and discrimination: A selective review and a connectionist model. Psychological Review, 101(4), 587–607. https://doi.org/10.1037/0033-295X.101.4.587
- Pearce, J. M., & Hall, G. (1980). A model for Pavlovian learning: Variations in the effectiveness of conditioned but not of unconditioned stimuli. Psychological Review, 87(6), 532–552. https://doi.org/10.1037/0033-295X.87.6.532
- Qian, L., Burrell, M., Hennig, J. A., Matias, S., Murthy, V. N., Gershman, S. J., & Uchida, N. (2025). Prospective contingency explains behavior and dopamine signals during associative learning. Nature Neuroscience, 28(6), 1280–1292. https://doi.org/10.1038/s41593-025-01915-4
- Rescorla, R. A. (1968). Probability of shock in the presence and absence of CS in fear conditioning. Journal of Comparative and Physiological Psychology, 66(1), 1–5. https://doi.org/10.1037/h0025984
- Rescorla, R. A. (1969). Conditioned inhibition of fear resulting from negative CS-US contingencies. Journal of Comparative and Physiological Psychology, 67(4), 504–509. https://doi.org/10.1037/h0027313
- Rescorla, R. A. (1988). Pavlovian conditioning: It's not what you think it is. American Psychologist, 43(3), 151–160. https://doi.org/10.1037/0003-066X.43.3.151
- Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy (Eds.), Classical conditioning II: Current research and theory (pp. 64–99). Appleton-Century-Crofts. https://www.columbia.edu/~rk566/Session4/Theory%20of%20Pavlovian%20Conditioning.pdf
- Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593
- Vervliet, B., Craske, M. G., & Hermans, D. (2013). Fear extinction and relapse: State of the art. Annual Review of Clinical Psychology, 9, 215–248. https://doi.org/10.1146/annurev-clinpsy-050212-185542
- Wagner, A. R., Logan, F. A., Haberlandt, K., & Price, T. (1968). Stimulus selection in animal discrimination learning. Journal of Experimental Psychology, 76(2), 171–180. https://doi.org/10.1037/h0025414