RESEARCH DOSSIER
Generalization, Discrimination, and Transfer of Learning
Abstract. Generalization, discrimination, and transfer of learning explain how acquired knowledge and behavior extend to, differentiate among, and remain effective in situations beyond original training. This fundamental research monograph integrates stimulus-control experiments, perceptual and category learning, exemplar and Bayesian models, associative prediction, contextual retrieval, structural analogy, and classroom transfer. It examines the historical transition from Thorndike and Woodworth's training-specific experiments through generalization gradients and modern theories of relational abstraction. Original experimental findings are evaluated alongside methodological reviews, recent integrative syntheses, and meta-analyses of cognitive training. The analysis distinguishes acquisition from immediate performance, near from far transfer, and supported task success from autonomous application. It identifies the conditions under which similarity, structured knowledge, discrimination, context, and practice schedule contribute to successful transfer, while reviewing evidence that many broad cognitive-training claims exceed the demonstrated outcomes. The final synthesis defines observable transfer tests and a research agenda for education, comparative psychology, and human–artificial learning environments.
Keywords: stimulus generalization; discrimination learning; transfer of learning; near transfer; far transfer; generalization gradient; categorization; psychological similarity; analogy; schema abstraction; context-dependent retrieval; interleaved practice; cognitive training; educational psychology; human–artificial learning
The Scientific Problem of Generalization
Generalization is the application of a learned response, prediction, category, or procedure beyond the particular observations on which it was acquired. Discrimination is the differentiation of responses when stimuli, situations, or consequences warrant different treatment. Transfer of learning describes the effect of prior experience on performance or acquisition in a new task or setting. These phenomena are jointly important because an adaptive learner must exploit regularities while recognizing exceptions. Their study examines how similarity is represented, which features guide selection, how environmental structure is inferred, and when what was learned remains usable after cues or demands change. The proposed synthesis treats transfer as an empirical outcome with specified training and test conditions, following the taxonomic argument of Barnett & Ceci (2002).
The relevant unit of analysis extends from sensory features to relations among abstract variables. A person who recognizes an unfamiliar member of a learned category demonstrates item-level generalization; one who applies a mathematical procedure to a structurally related word problem may demonstrate relational transfer. In contrast, practicing a single computer task to higher speed establishes within-task acquisition, while improvement on a task with different material and response requirements requires an independent test. Generalization can sometimes be inappropriate, as when an acquired fear response spreads to safe signals, and discrimination can sometimes be too narrow, as when an individual fails to apply a principle outside its classroom example. These functions are evaluated against the organism's actual environment and goals.
A coherent scientific account must explain at least five operations: encoding the training experience; representing relevant commonalities and differences; selecting or retrieving a prior response; adapting the response to the new situation; and monitoring the consequences. The relative influence of these operations changes across conditioning, categorization, motor expertise, education, and social interaction. The 2025 synthesis by Wu et al. (2025) identifies the continuing dialogue between rule-based, similarity-based, and hybrid accounts. Each explains part of the evidence. The research program therefore requires separate predictions about gradient shape, categorical boundaries, transfer distance, retrieval cues, and individual adaptation rather than treating the general word generalization as an explanation in itself.
Constructs, Operational Definitions, and Boundary Conditions
Stimulus generalization is commonly operationalized by measuring response probability, frequency, latency, or magnitude to a range of untrained stimuli after controlled acquisition. A generalization gradient plots the response against systematically varied stimulus values or psychological distances. A discrimination procedure establishes differential outcomes for two or more stimulus classes and examines whether responding comes under their control. Transfer, by comparison, is operationalized through the effect of a training condition on performance in a criterion task not simply identical to practice. The dependent variable may be accuracy, speed, retention, error type, strategy selection, or achievement in a real environment. Each measure answers a different question; a steep gradient and successful cross-task transfer need not covary.
Near and far transfer specify degrees of change in task, material, function, setting, or required response. They should be stated operationally instead of inferred from an investigator's intuition about superficial similarity. Barnett & Ceci (2002) argues for characterizing transfer across multiple dimensions, including the content learned and the context of application. Transfer may be positive when prior learning improves later performance, negative when earlier responses interfere with a new task, or neutral in a defined outcome even after demonstrable acquisition. These outcomes depend on baseline ability, task difficulty, opportunity to apply learned knowledge, and whether a newly acquired schema is actually called upon during testing.
Important boundaries separate generalization from statistical generalizability and from developmental general growth. A learning experiment concerns whether trained information affects a new stimulus or task for its participants; external validity concerns whether findings extend across samples, institutions, and populations. Both require careful design, but they answer different causal questions. Likewise, successful categorization of new examples is a stronger inference about representation than retention of the original examples, yet it does not by itself prove broad reasoning improvement. The scientific objective is to identify the locus and conditions of changed performance. Strong conclusions specify the training manipulation, novel test domain, delay, comparator, and plausible mechanisms.
Early Transfer Research and the Specificity of Practice
Early experimental psychology was deeply concerned with the educational belief that exercising a general mental faculty could improve performance across diverse activities. Edward Thorndike and Robert Woodworth investigated whether improvement in one function transferred to others, using pretraining and posttraining measures on neighboring tasks. Their 1901 investigations exposed the central difficulty of treating superficially related activities as one homogeneous faculty. The original historical report Thorndike & Woodworth (1901) describes variations in mental operations, training tasks, and criteria of improvement. Its logic remains recognizable in contemporary controlled transfer designs: establish change on the trained activity, define an independently measured target, and ask what components or relations truly carry over.
The intellectual consequence of this early work was a shift toward specificity of training and analysis of shared elements. A change might transfer because two tasks recruit common stimulus discriminations, response sequences, conceptual relations, or strategies. Transfer then becomes a question about component overlap and functional invariance. However, literal physical overlap is not the sole possible route. Structural analogy can permit use of a solution learned in one story or scientific domain in an apparently different problem. The history thus contains an enduring tension between experience tied to detailed contexts and abstraction that permits flexible reuse. Neither tradition entails that every form of transfer is always weak or always strong.
Harlow's influential work on learning sets provided a further historical route to cross-problem improvement. In repeated discrimination problems, an organism may improve at selecting an effective strategy for a new discrimination even when individual stimuli differ. The article Harlow (1949) established the term learning set in a controlled comparative learning tradition. The phenomenon raises a question that cannot be answered by simple item retention: whether successive tasks change how new tasks are approached. The precise generality of a learning set must still be established by new stimuli, task types, delays, and comparisons. From this point forward, learning research had to distinguish specific associations, learned procedures, and higher-order regularities in task structure.
Stimulus Generalization Gradients
Classical stimulus-control experiments reveal how learned responding spreads across a continuum. Following differential or nondifferential training with a particular stimulus, investigators introduce test values along one dimension and observe response changes. Guttman and Kalish's 1956 investigation Guttman & Kalish (1956) is a foundational study of discriminability and stimulus generalization. Such experiments transformed the broad idea of similarity into a measurable relation between training history and test responses. The interpretation requires care: a gradient is a behavioral function under a particular testing protocol, rather than a direct picture of a stored representation. Response extinction during testing, general arousal, and perceptual acuity can all shape the observed curve.
One central question is whether distance should be measured in physical stimulus units or in psychological discriminability. Two colors separated by an equal wavelength interval may differ more or less in perceived similarity depending on background, category structure, and sensory system. Shepard (1987) derived an approximately exponential law for generalization as a function of distance in an appropriately constructed psychological space, subject to the assumptions of the model. This account relates observable probability to representational geometry and motivates quantitative comparisons between candidate distance metrics. It also makes clear why simple physical similarity is insufficient as a general theory of learning: the geometry of perceived difference may itself change with experience and attention.
Gradients can be broad because distinctions are difficult, because outcomes are uncertain, or because a learned category is represented at an abstract level. Conversely, steep gradients can reflect sharp sensory discriminability, highly selective attention, or explicit rules. These alternatives imply different intervention tests. Manipulating perceptual spacing while keeping reinforcement constant tests one set of explanations; changing consequences while preserving perceptual features tests another. A credible gradient experiment samples sufficient novel stimuli, varies test order, and accounts for serial carryover. The empirical curve must be considered alongside the task's reinforcement history and the participant's opportunity to infer a dimension or category rule.
Discrimination Training, Peak Shift, and Inhibitory Control
Discrimination learning produces different treatment of events that originally appeared similar. In a basic design, one stimulus signals reinforcement or a particular outcome, while a neighboring stimulus signals another outcome or its absence. Subsequent tests may show responding concentrated around the reinforced stimulus, with suppression around the comparison stimulus. Hanson's 1959 study Hanson (1959) is a landmark experimental investigation of how discrimination training alters generalization. It is historically associated with peak shift, in which the highest response on the gradient can move away from the explicitly nonreinforced comparison. The exact outcome depends on stimulus dimensions and training conditions, not on an invariant rule that all discrimination produces such a shift.
Peak shift can be interpreted through interaction between excitatory and inhibitory stimulus gradients, attention to relational dimensions, or changed representations of the stimulus space. These explanations predict different effects from manipulating test distance, response criterion, training intensity, and contextual cues. The crucial empirical finding is that the shape of posttraining generalization is a property of learning history. Responses to a novel stimulus need not reflect its raw similarity to an original positive example. They can reflect what contrasts were emphasized, which errors were punished, and how categories were operationally defined. Thus a learned distinction can reorganize the field of possible future reactions.
Discrimination is also a prerequisite for useful generalization. A novice may respond similarly to many cases because relevant dimensions are unavailable; training can sharpen the detection of those dimensions. Overtraining a narrow boundary, however, can undermine flexible action when task priorities change. Evidence about improved sensitivity must separate true perceptual change from criterion changes. In experimental reports it is useful to model sensitivity and decision threshold separately where the task permits. Discrimination training should be evaluated by whether it improves responses to valid distinctions under new test conditions, including difficult boundary items and unexpected combinations of features.
Perceptual Learning and the Modification of Similarity
Similarity is partly constructed by the learning process under investigation. Category training can alter attention to diagnostic dimensions, changing which differences carry weight in later identification and discrimination. Nosofsky's exemplar framework Nosofsky (1986) connects identification data, categorization judgments, and selective attention in a quantitative theory. Its significance extends beyond any single stimulus set: a learner may compare new items with remembered instances in a psychological space whose dimensions have been reweighted by experience. Such a model can represent graded membership and exceptions while making predictions about confusion patterns. Exemplar retrieval is therefore a mechanistic alternative to a single fixed prototype or explicit decision rule.
Goldstone's experimental study Goldstone (1994) examined how categorization regimes affected subsequent perceptual discrimination. In the reported experiments, category training yielded evidence of acquired distinctiveness, with stronger sensitivity to certain between-category differences, and a more restricted instance of acquired equivalence. These results support the possibility that classification experience can change later perceptual judgments. Their interpretation nonetheless demands attention to task instructions, response criteria, stimulus dimensions, and the durability of the observed effects. Not every change in categorical choices indicates a transformation of low-level sensory processing; attention, comparison policy, and memory can also contribute to the measured outcome.
A broader implication follows for transfer. The degree of similarity between a trained and an untrained task is not always fixed in advance: expertise may reorganize the space in which transfer distance is evaluated. The learned salience of a feature can make previously separate examples seem alike, and previously similar examples functionally different. This creates reciprocal causality between discrimination and generalization. The environment supplies contrasts, learning changes the representation of those contrasts, and modified representations govern responses to subsequent cases. Such reciprocity can be examined experimentally by measuring both identification and category use before and after training, with appropriate independent testing and controls.
Exemplar, Prototype, and Rule-Based Accounts
Categorization models disagree about the representational unit that licenses generalization. Exemplar theories retain information about encountered instances and compute the relation of a new item to stored examples; prototype theories emphasize central tendencies; rule-based models classify by explicit or inferred feature relations. Medin and Schaffer's context model Medin & Schaffer (1978) demonstrated the explanatory value of exemplar retrieval for ill-defined categories and challenged accounts that assume purely additive feature information. The practical importance is that a new observation can be classified differently depending on whether it resembles a familiar exception, matches an average pattern, or satisfies a learned conditional rule. A transfer design can distinguish these accounts by manipulating diagnostic feature combinations and withholding critical examples.
Different category organizations demand different forms of representation. A simple threshold on a single relevant variable may support an explicit decision rule and rapid transfer across superficial changes. A family-resemblance category with overlapping but nondefining cues may favor similarity-weighted classification. A category organized by causal relations or functional roles may require a representation of how features interact. Models should therefore be compared on predictions for transfer stimuli rather than only on goodness of fit to training cases. Flexible complexity can fit many observations, so model comparison benefits from held-out predictions, parameter checks, and experiments designed to make explanations diverge.
Theoretical progress requires separating the information learned from the process that expresses it. A learner may acquire multiple examples and deploy them through a selective-attention rule; another may use explicit language to describe a rule while relying on familiarity in a speeded task. These possibilities are compatible with some training results and distinguishable with transfer tests that vary response time, load, instruction, and exception frequency. A general law at the level of responses remains valuable, but it need not imply a single cognitive algorithm. The joint study of categorical generalization and discrimination provides an unusually clear setting for examining the explanatory trade-offs among representations.
Bayesian Generalization and Latent Structure
Bayesian accounts begin from uncertainty about the source of observations and the structure generating them. They estimate how likely a new item or context is to belong to a learned class under competing hypotheses. Tenenbaum and Griffiths Tenenbaum & Griffiths (2001) extended the treatment of similarity-based generalization by deriving predictions from assumptions about sampled examples, natural kinds, and hypothesis spaces. This framework explains why identical perceptual distances can imply different expectations when the learner considers distinct category boundaries or causal structures. Generalization depends not only on the closeness of examples but also on beliefs concerning why those examples appeared.
The theory is especially informative when sparse observations support multiple plausible rules. One encountered animal might represent a biological species, an ecological group, or a broader category; the inferred scope of a learned property changes with the hypothesis that organizes the evidence. Additional examples can reduce uncertainty or shift attention to a different latent partition. The resulting learning may be discontinuous at the level of declared categories even when sensory distances change smoothly. A model of generalization should therefore state its assumptions about sampling, prior probability, feature dependence, and the learner's access to abstract relational descriptions.
Contemporary integrative reviews describe hybrid approaches that combine structured inference with similarity-based approximation. Wu et al. (2025) places concept learning, function learning, and reinforcement learning within a shared history of the generalization problem. These approaches can be compared on held-out tests involving extrapolation, rule switches, ambiguous category boundaries, and changes in task reward. A Bayesian model is an explanatory proposal whose validity depends on assumptions and predictive accuracy; the mere possibility of rational inference does not establish that every human response approximates the ideal observer. Systematic individual differences, processing costs, and instructional history constrain the usefulness of normative comparisons.
Reinforcement Learning and Generalization of Value
In associative and reinforcement learning, prior experience guides predictions about future consequences. A cue can acquire a response tendency when it predicts a meaningful outcome, and a related cue may evoke a similar tendency without direct training. Simple error-correction models update an expected value following discrepancy between prediction and observation. This mechanism can account for acquisition under controlled contingencies but leaves open how similarities among untrained cues are represented. Associative models often require assumptions about stimulus coding, attention, or context to yield specific generalization gradients. The generalization problem therefore reappears inside the computational representation: which experiences share parameters, features, or inferred latent causes?
Gershman's probabilistic analysis Gershman (2015) demonstrates how accounts of uncertainty and reinforcement updating can be brought into one framework rather than treated as wholly separate explanatory languages. It makes the representation of hidden causes and the treatment of uncertainty central to learning. A learner can generalize because different observations are assigned to one cause, and discriminate because evidence supports separate causes or contingencies. In human experimental tasks, these possibilities are studied by changing transition statistics, reward functions, or the salience of latent contexts and then testing decisions with new cue combinations. Model comparisons must examine actual trial-by-trial predictions as well as aggregate posttraining accuracy.
Value generalization can be either advantageous or costly. A previously rewarded strategy may be applied efficiently to a novel situation with the same causal structure, yet it may persist when outcome contingencies have changed. Reversal training and contextual shifts expose whether the learner relies on a cached value, an explicit rule, or an inferred state representation. These distinctions carry into education and professional practice, where a fast familiar procedure can compete with a more accurate interpretation of a new problem. Good experiments establish whether transfer derives from shared representations, retrieval of a known response, or new learning during the test itself.
Context as a Gate on Generalization and Retrieval
Context includes the physical environment, preceding events, internal state, task framing, and temporal circumstances that make a particular response meaningful. The same stimulus can acquire different consequences under different contextual arrangements. Bouton's analysis Bouton (1993) argues that contextual cues influence retrieval when learning phases supply conflicting information, such as acquisition followed by extinction. This perspective shows why a response can reappear after a setting changes, even when training appeared successful in the original setting. A context switch may alter behavioral expression without reversing the underlying acquisition history. Consequently, broad claims about retained or extinguished learning require tests beyond the immediate training context.
Contextual effects have several candidate mechanisms. Features of a new environment may function as direct predictors, disambiguating signals, retrieval cues, or components of a configural representation. Training in varied environments can reveal whether performance depends on specific cues or on generalizable relationships. Transfer failures must therefore be interpreted using evidence from both memory probes and behavioral measures. The crucial distinction is between a representation that was never acquired, one that cannot be retrieved under new cues, and one that is retrieved but cannot be appropriately adapted. These explanations predict different outcomes under reminders, explicit instructions, or changing contextual features.
Time itself changes the availability of retrieval cues and intervening knowledge. A delayed transfer test is stronger than immediate performance as evidence of enduring applicability, yet it can also introduce new experiences and interference. Researchers should vary delay independently from contextual similarity where possible and record whether the test includes new feedback. Feedback may generate learning during assessment, changing the dependent variable from generalization of the original training to the joint effect of training and test exposure. Context dependence is thus a central theoretical topic and an experimental design constraint rather than a nuisance variable to be removed by routine averaging.
Symbolic Relations, Fear Generalization, and Safety Discrimination
Generalization may follow inferred symbolic relations even when stimuli look unlike one another. A learned threat signal can affect responses to cues related through category membership, language, or trained equivalence rather than simple sensory resemblance. A systematic review by Boldrin et al. (2025) covers human work on symbolic or derived generalization of fear and avoidance, drawing attention to the role of learned relational networks. These results enrich the concept of generalization by showing how functions of events may spread through acquired relations. Interpretation demands precise tasks: the strength of a derived response, the instructional history, and demonstrated mastery of the relation all influence what a result can support.
Fear-conditioning experiments require an especially careful distinction between threat learning and discrimination of safety. Physiological arousal, expectancy ratings, avoidance choices, and verbal judgments can index partially different processes. A broad gradient of defensive responding may indicate difficulty identifying the safety boundary, broad threat expectation, or high response sensitivity. It does not independently establish a clinical diagnosis. Studies comparing conditioned responses to trained danger cues, safe cues, and generalization stimuli can examine whether responding tracks physical similarity, symbolic relations, or the inferred probability of an adverse event. Independent replication and control of demand characteristics are essential before making clinical inferences.
The clinical relevance of the phenomenon lies in the possibility that harmless cues acquire defensive functions and that safety learning is not reliably retrieved in altered contexts. However, mechanistic laboratory demonstrations do not automatically specify the effectiveness of a therapeutic intervention. Differences between acute experimental conditioning, chronic symptoms, treatment context, and individual life histories remain substantial. The correct translational pathway proceeds from operationalized mechanisms to clinical samples, controlled interventions, and longitudinal outcomes. The same generalization and discrimination vocabulary can connect basic and clinical research while preserving the distinct evidential requirements of each.
Relational Representation and Analogical Mapping
Transfer across domains often depends on recognizing a common relation rather than a shared surface feature. A drainage network and an electrical circuit may look different while permitting structurally related analyses of flow, resistance, and constraint. Structure-mapping theory Gentner (1983) describes analogy as the alignment of relational systems, with correspondences among entities constrained by the network of relations. This account specifies how a learner could move beyond literal resemblance. The important empirical claim concerns whether emphasizing relational structure improves later problem solving in a different domain, and under which cues the source example becomes accessible. Relational transfer is therefore distinct from simply naming two cases as similar.
Analogical transfer requires multiple processing stages that may fail independently. A relevant source must be retrieved, mapped to the target, evaluated for applicability, and adapted to the target's details. Retrieval often favors surface cues because those cues readily activate familiar episodes; successful mapping, however, depends on deeper structural correspondence. Holyoak and Koh's work Holyoak & Koh (1987) examined the respective roles of surface and structural similarity in analogical transfer. The resulting distinction is essential for designing tests: a novel problem can resemble training materials without sharing the causal solution, and another can demand the same solution despite a different narrative. Both kinds of probes are needed.
A strong analogy is a constrained hypothesis about relational organization. It may identify an informative correspondence but still fail because the target contains a boundary condition absent from the source. Instruction should therefore teach comparison of relational structure together with evaluation of where the analogy breaks. The developmental ability to use relational comparisons also depends on knowledge, available working-memory resources, and the complexity of the mapping. Across disciplines, transfer research must specify whether the learned object is a concrete answer, a reusable schema, a causal model, or a strategy for recognizing which problem structure is present.
Classic Experimental Evidence for Analogical Problem Solving
Gick and Holyoak's investigations Gick & Holyoak (1980) used a source narrative and a target problem to examine the conditions under which people retrieve a relevant analogy. Their original experiments showed that a solution from a military story could inform a medical problem when participants were prompted to use the prior story. This design established an influential separation between possessing a useful source example and spontaneously recognizing its relevance. An unsuccessful unaided transfer attempt does not prove that the learner lacks a transferable representation: retrieval and selection can be the immediate obstacle. Conversely, transfer after an explicit hint does not demonstrate spontaneous transfer under ordinary conditions.
The comparison of spontaneous and cued performance is methodologically decisive. A prompt can reduce search demands, clarify that analogical reuse is expected, and change the person's interpretation of the task. It thereby helps isolate whether mapping and adaptation are possible once the source is activated. The experiments also manipulated aspects of analogical fit, revealing that mere exposure to a previously successful solution is insufficient. This approach supplies an experimental vocabulary for later instructional designs: source availability, retrieval cue, mapping quality, adaptation, and target success can be measured separately rather than collapsed into a single transfer score.
The ecological challenge is that classroom and workplace tasks rarely announce which earlier example is relevant. Professionals face competing precedents, ambiguous goals, and constraints that require selection among possible analogies. Laboratory source–target studies provide a controlled way to isolate a mechanism, but field outcomes depend on task prevalence, time pressure, and incentives for trying a transferred strategy. A complete account must connect the experiment's internal validity to realistic environments where the learner must identify the opportunity to transfer and judge whether doing so is sensible.
Comparison-Based Learning and Schema Abstraction
Placing two examples side by side can make shared relational structure more visible. Gentner, Loewenstein, and Thompson Gentner et al. (2003) tested analogical encoding in studies of negotiation knowledge, comparing the instructional consequences of learning from cases and directly comparing their structure. The experiments support the idea that structured comparison can improve the abstraction and later use of relevant strategies under the tested conditions. The active ingredient is not merely seeing more cases; it is engaging in alignment that highlights a common principle and clarifies its scope. This mechanism explains why additional examples alone may have less impact than a smaller set deliberately organized for contrast.
Comparison can promote discrimination at the same time as generalization. When cases share a structural solution but differ in surface detail, aligning them encourages abstraction from irrelevant features. When they differ at a decisive causal step, juxtaposition can reveal a boundary that prevents overgeneralization. This dual function is important in science education, diagnostic reasoning, and legal analysis. Effective instruction requires examples that reveal both invariants and exceptions, followed by tests using new materials. The transfer hypothesis is strongest when the novel tasks differ systematically from taught cases and when outcome measures reward accurate selection rather than rote reproduction.
Comparison interventions also have cognitive costs. If learners lack prerequisite knowledge, the effort of aligning complex cases can overload available resources, leading to superficial matching. Guidance may initially be needed to make relations explicit, but its eventual removal is part of testing independent competence. An intervention should therefore examine learner knowledge, quality of explanation, delayed retention, and autonomous choice of analogies. What is transferable may be a problem schema, a strategy for generating comparisons, or both; distinct assessments are needed to disentangle these outcomes.
Near Transfer, Far Transfer, and the Multidimensional Taxonomy
The everyday distinction between near and far transfer is useful for orienting a research question but too coarse to serve as a complete measurement system. Barnett & Ceci (2002) demonstrates that transfer can differ across knowledge domain, physical and social context, temporal separation, functional use, and other dimensions. Two investigators may use the same phrase far transfer while measuring very different departures from training. A laboratory classification task with new pictures might be near in response demands but far in stimulus appearance; a decision in a new workplace might retain content yet change incentives, social roles, and required actions. Valid comparison requires describing the vector of changes rather than only its presumed distance.
The transfer criterion must be declared before the result is interpreted. Is the target a new example of the same kind, a new configuration of familiar elements, a novel rule with shared structure, or a remote real-world outcome? Does the test involve the same instructions, equipment, assessor, or feedback? Are participants aware that the source is applicable? These details alter the causal mechanisms involved. Moreover, transfer can refer to immediate reuse, accelerated learning of a later task, or durable independent performance after an interval. Each should be tested through procedures that make the target's novelty and the transfer opportunity observable.
Pooling disparate transfer outcomes in one average can conceal genuine moderator structure and create artificial disagreement. For example, weak transfer to general intelligence after narrowly targeted practice does not imply weak transfer to a close reasoning variant; strong transfer between analogous classroom problems does not imply improvement across every domain of life. The taxonomic approach makes these findings compatible without reducing them to a vague universal conclusion. A science of transfer advances by predicting which dimensions of difference are tolerated by which learned representations and by testing such predictions across independent populations and task families.
Retrieval, Encoding Specificity, and Task Demands
Acquired knowledge becomes useful in a new setting when the learner can retrieve and deploy it under current cues. Morris, Bransford, and Franks Morris et al. (1977) experimentally contrasted kinds of encoding and recognition tests, demonstrating that the advantage of a study operation depends on the later task. Their findings helped establish transfer-appropriate processing as an organizing principle: performance depends partly on whether operations practiced during encoding are relevant to the conditions of retrieval. This is distinct from claiming that any match guarantees learning. Study activities can support durable semantic understanding, yet tests emphasizing other operations may expose different strengths and weaknesses.
Transfer-appropriate processing creates a design trade-off. Practicing exactly the later test can raise performance without building flexibility, whereas practicing an unrelated operation can leave the learner unprepared. A defensible instructional solution is to identify the underlying operations likely to be needed and vary their surface conditions. Retrieval from multiple cues may create more routes for accessing prior knowledge, and explanatory practice may support flexible reconstruction when direct recall fails. These are hypotheses to be tested under independent target conditions; apparent benefits during guided practice are only preliminary evidence of transfer.
Task demands also determine whether a learner recognizes an opportunity to use what was learned. A change from free recall to recognition, from multiple-choice questions to explanation, or from guided exercises to independent action can alter the observed outcome. To interpret transfer, investigators should report whether training and testing require comparable response production and whether any difference was intentionally manipulated. When the goal is independent competence, the final test should exclude the supports that originally supplied essential intermediate reasoning, while measuring the learner's ability to reconstruct a solution.
Practice Conditions and the Difference Between Acquisition and Retention
High accuracy during a training session may reflect immediate repetition, familiar cues, and the availability of support rather than knowledge capable of surviving delay or contextual change. Schmidt and Bjork Schmidt & Bjork (1992) synthesized evidence from motor and verbal learning to argue that procedures optimizing short-term practice performance need not optimize retention and transfer. This distinction challenges the assumption that easier, faster learning sessions necessarily produce more broadly usable skills. An empirical transfer program should therefore assess acquisition curves, delayed retention, and performance on changed tasks as separate outcomes. The difference between these variables can itself reveal how the training structure shaped representation and retrieval.
Variable practice exposes the learner to different parameter combinations, perceptual inputs, or task conditions. It can strengthen identification of the invariant relation required for success, but it can also increase errors when examples are too complex or poorly sequenced. Interleaving multiple task categories forces selection of a suitable procedure rather than application of a recently demonstrated one. The hypothesized transfer benefit concerns discrimination among problem types and selection of strategies under novel conditions. Whether variation improves learning depends on the learner's prior knowledge, availability of feedback, and similarity of training and criterion tasks.
A rigorous design matches groups on instructional time and relevant exposure while varying the organization of practice. A single posttest is insufficient to determine whether a method affected enduring generalization. The test should include practiced items, genuinely novel examples, and tasks that require flexible discrimination among strategies, ideally after a meaningful interval. The measurement plan should also identify whether participants received feedback during the transfer test. A procedure that appears difficult during practice may support better delayed performance, but the effect must be shown for the particular domain and learner group rather than assumed from a general slogan about desirable difficulty.
Interleaving, Spacing, and Category Induction
Learning to discriminate among closely related categories often benefits from opportunities to compare multiple classes. Kornell and Bjork Kornell & Bjork (2008) tested how the presentation order of paintings influenced later induction of artists' styles. In their experiments, spaced presentations supported induction even when learners' judgments of the easier massed condition were more favorable. The result illustrates a dissociation between experienced fluency and the capacity to categorize novel examples. Its theoretical interpretation concerns comparison and attention to diagnostic differences. It does not imply that every subject, age group, or classification problem benefits from the same degree of interleaving.
In mathematics practice, Rohrer and Taylor Rohrer & Taylor (2007) distinguished distributing repetitions from mixing problem types. Their experiments reported advantages on delayed tests for the studied spacing and interleaving arrangements, supporting the value of practice schedules that require remembering and choosing procedures. The mechanisms may differ: spacing changes retrieval demands across time, while interleaving increases the need to discriminate which strategy applies. An instructional intervention claiming broad transfer must specify which dimension it manipulated and whether the outcome required use on a new kind of mathematical problem rather than one already rehearsed in altered order.
The general lesson is to organize examples so the learner can identify both similarities worth reusing and distinctions that determine when a strategy applies. Interleaving is particularly plausible when alternatives are confusable and when successful performance depends on correct method selection. Blocked practice may retain advantages for initial acquisition of an unfamiliar procedure or when cognitive demands are excessive. Current evidence supports conditional recommendations: test delayed and novel-task outcomes, vary prior knowledge, and examine whether feedback permits learners to resolve confusion. Training procedures must be judged by their observable transfer outcomes, not by the appealing appearance of their schedules.
Instruction for Transfer in School and University Settings
Educational transfer demands the coordinated development of domain knowledge, recognition of relevant structure, and the ability to adapt a solution under new task demands. Teaching a reusable formula is valuable only if learners can recognize when it applies, explain why, and reject it when its assumptions are violated. A lesson can make common structure visible through comparison, ask students to predict contrasting outcomes, require explanation of conditions, and then remove cues in independent assessments. The results of Gentner et al. (2003) support focused comparison as one route to relational learning. The broader implementation should be evaluated against authentic instructional constraints and not simply inferred from a laboratory task.
Nokes-Malach and Mestre Nokes-Malach & Mestre (2013) describe transfer as sense-making in which learners represent a target situation, generate possible solutions, and evaluate whether a prior schema fits. This perspective connects cognitive transfer to goal interpretation, prior knowledge, and metacognitive judgment. It also clarifies why students who can recite a principle may fail to deploy it: they may identify the new task differently or lack the resources to map the known solution onto its constraints. Assessment should sample spontaneous recognition of opportunities, explicit justification, and successful execution. Each outcome identifies a different instructional achievement.
Instructional experiments face a persistent tension between standardized tasks and real educational consequences. Tasks that are too similar to examples may inflate transfer claims; tasks that change every relevant dimension may test many abilities besides the learned target. Curriculum studies must define appropriate distance, document teachers' implementation, and account for prior achievement and classroom clustering. Strong practical evidence combines preregistered comparisons, multiple outcome measures, delayed tests, and replication in different schools. The goal is a precise account of which teaching method improves which independent skill for which learners, rather than a promise that one practice universally improves every intellectual capacity.
What Cognitive Training Meta-Analyses Establish
Cognitive training has generated prominent claims that practice on narrow executive or working-memory tasks improves general intelligence, school achievement, or daily functioning. The crucial evidential question is whether the trained group improves on independent abilities more than a credible comparison group, beyond expectancy, repeated testing, and task familiarity. The meta-analysis by Melby-Lervåg et al. (2016) examined working-memory training with attention to treated control groups. It reported immediate improvements on some related memory tasks but found no convincing reliable far-transfer benefits for the studied broad outcomes when comparisons used treated controls. This is evidence about a specific family of training programs and tests, not a universal prohibition on all forms of learning transfer.
A broad assessment of commercial and research brain-training claims by Simons et al. (2016) likewise emphasized that improvements on practiced tasks and related activities do not establish generalized cognitive benefits. The review evaluated methodological quality, theoretical plausibility, and evidence for applications beyond training. Sala & Gobet (2019) surveyed meta-analytic evidence across several cognitive-training domains and argued that general cognitive improvement is minimal and sensitive to design quality. Such syntheses require scrutiny of study overlap, dependence among outcomes, publication bias, and the definition of active comparison conditions. Their central contribution is to distinguish improvements that a program reliably trains from distant capacities it is marketed as changing.
Negative or small average effects should be interpreted at the level of the measured intervention and criterion. Some near-transfer outcomes may be positive, and important moderators may remain unresolved. Nevertheless, a program promising broad transfer must demonstrate it through appropriately designed trials, rather than relying on biological plausibility, improvements on a training score, or correlational associations between experts and novices. Research should test whether genuine common processes exist between the training task and distant criterion, whether they are modified by intervention, and whether their modification mediates the target benefit. This is a more discriminating standard than declaring every transfer either present or absent.
Field Evidence and the Limits of Classroom Generalization
Field research offers essential tests of whether controlled learning principles extend to institutional settings. A 2025 analysis by Jerrim and colleagues Jerrim et al. (2025) examined associations between reported mathematics teaching practices and unfamiliar problem performance using TIMSS data. The study did not find evidence that the examined schema-oriented practices were associated with improved performance on its particular near-transfer criterion. This is an informative constraint on broad pedagogical expectations, but its observational design does not supply the same causal identification as randomized assignment of teaching methods. Reported instruction, policy context, curriculum, and measurement can introduce alternative explanations.
The contrast between successful laboratory analogical-encoding experiments and weaker large-scale classroom associations highlights the importance of matching theory to implementation. A comparison prompt in a tightly controlled task may create the intended relational operation, while a teacher-report variable in a survey may only imperfectly approximate that operation. Classroom success also depends on consistent instruction, opportunity to practice, and the fit of the criterion test to taught knowledge. The right response to discrepant evidence is to develop studies with direct process measures and well-defined interventions, not to treat one design as automatically overriding all others.
A strong educational evidence program combines experiments that isolate a mechanism, well-powered classroom trials, and follow-up observational research into generalizability. It distinguishes student-level changes from school-level composition and accounts for clustered assignment. The research must record which examples were taught, what cues the teacher supplied, how much independent problem solving occurred, and whether the outcome was scored blind to condition. Durable transfer is demonstrated by successful use after scaffolding is absent and the learner encounters a genuinely altered problem. These requirements bring generalization research from the stimulus continuum into educational practice without relaxing its experimental logic.
Development, Expertise, and Individual Differences
Generalization changes with the learner's knowledge, developmental history, sensory capacities, and experience in a domain. Children may have difficulty using an abstract relation when its constituent concepts are unfamiliar; an expert can recognize the same relation quickly because a rich knowledge structure supports interpretation. These differences alter what qualifies as a distant task. For an experienced mathematician, two superficially different problems may instantiate a familiar schema; for a novice they may require independent reasoning operations. Developmental assessment must therefore compare structurally matched tasks at appropriate difficulty levels rather than equating chronological age with a single generalization capacity.
Diamond and Lee's review Diamond & Lee (2011) examines interventions associated with development of executive functions in children and illustrates the diversity of training approaches and outcomes. The relevance for transfer research lies in separating improvement on particular executive-function tasks from broader educational, behavioral, or developmental consequences. Schooling itself has wider effects: a quasi-experimental meta-analysis by Ritchie & Tucker-Drob (2018) examined educational influences on cognitive test performance using designs intended to reduce confounding from prior ability. Such findings concern sustained educational exposure and broad developmental opportunity, which should be analyzed separately from the far-transfer claims of a brief computerized cognitive drill.
Individual differences also affect the probability that transfer will be expressed. Motivation, task interpretation, prior knowledge, language proficiency, anxiety, and metacognitive calibration can change whether a learned response is retrieved and applied. A participant who possesses a principle but doubts its relevance may fail a spontaneous transfer test; another may guess correctly without genuinely understanding the relation. Researchers can probe explanations, confidence, and strategy selection alongside behavioral success. Population-level findings should therefore be accompanied by analysis of measurement invariance, meaningful subgroup variation, and whether developmental and cultural experiences shape what the test actually demands.
Skills, Sensorimotor Generalization, and Cross-Domain Boundaries
The domain of motor learning provides a direct illustration of context- and parameter-specific transfer. Practice changes predictions about the relation between action commands, bodily state, and sensory consequences. A learned control policy may apply to unfamiliar movement amplitudes but fail when the dynamics or feedback structure changes sharply. The training-to-test relation can be manipulated through altered targets, limbs, tools, or environmental forces. As in conceptual learning, transfer requires an operational definition of what remains constant. Similar-looking movements can differ in control demands, while movements that look different can share an underlying coordination principle.
The broad practice analysis by Schmidt & Bjork (1992) connects motor and verbal learning precisely because both domains display dissociations between immediate performance and lasting transfer. Yet this convergence does not imply identical underlying neural mechanisms. In sensorimotor experiments, changes in error correction, calibration, feedback sensitivity, and explicit re-aiming may produce different generalization profiles. A task may reveal preserved response parameters and still require rapid online recalibration. For this reason investigators should distinguish retention of an acquired parameter from accelerated learning of a new one, and measure the contribution of explicit strategies when possible.
Cross-domain transfer is best understood as a hierarchy of claims. Reusing a movement policy at a nearby target is one type of generalization; applying an explicit rule in a different physical system is another; improving general executive ability through motor practice is a far broader hypothesis. Each claim has its own control conditions and outcome criteria. The scope of an intervention's effects cannot be inferred from the elegance of its mechanism. Multi-condition testing across several distances provides a more informative generalization profile than a single binary assessment of whether transfer occurred.
Human–Artificial Learning Environments and Independent Competence
Artificial systems increasingly provide worked solutions, explanations, corrections, adaptive prompts, and simulated examples to human learners. These interactions create a distinctive transfer question: does assisted success become a capacity that the human learner can execute independently? An external system may improve task performance by supplying missing intermediate steps, selecting strategies, or identifying relevant prior cases. The observable quality of the final artifact then reflects the combined system. Independent transfer requires an assessment in which assistance is removed or explicitly varied and the human participant confronts unfamiliar but theoretically related tasks. This design separates the performance of a partnership from learning retained by a participant.
Generalization under artificial support also has a representational dimension. A tutor can present multiple examples with invariant structure, pose diagnostic contrasts, or ask learners to explain why a retrieved solution applies. Such designs follow mechanisms studied in analogy and discrimination research, including Gentner et al. (2003) and Nokes-Malach & Mestre (2013). Their effectiveness remains an empirical matter. An explanation generated by a system may be understood, memorized, accepted uncritically, or ignored. Process measures should examine what the learner attends to, which deductions they can reconstruct, and which errors they can recognize without the original prompt sequence. Assistance can be deliberately withdrawn as part of the learning design.
Artificial Era research also introduces a second object of study: generalization by artificial learning systems themselves. Human learning and model generalization can be compared at the level of observable performance, distribution shift, compositional structure, and updating under feedback. The comparison requires explicit metrics and careful matching of task conditions. A model that handles held-out examples may still rely on features unavailable or irrelevant to humans, while a human may transfer by causal interpretation or social contextualization. These are researchable differences in functional organization. The philosophical horizon From Homo to Artificial motivates systematic comparison but supplies no substitute for independent evaluation of either kind of learner.
Measurement, Causal Identification, and Reproducible Transfer Tests
A defensible transfer claim requires at least a specified training condition, evidence of acquisition, an independently defined target, an appropriate comparison, and an interpretable outcome. Ideally investigators preregister the primary transfer distance and endpoint, determine the necessary sample size, and make stimulus sets and analysis code available. Random assignment supports causal inference when feasible; active comparison tasks help control expectations and engagement. Baseline measures allow adjustment for existing skills but do not remedy all biases. Investigators should also report attrition, selective exclusions, missing data, and whether outcomes were selected after inspecting the results. Evidence of a trained-task effect must not be mistaken for validation of the transfer hypothesis.
Multiple test items are required to generalize beyond one idiosyncratic problem. A study that samples many participants but only a handful of transfer stimuli may estimate participant variability while neglecting stimulus variability. Better designs treat individuals, items, and testing sites as relevant sources of uncertainty. Held-out tasks should vary along prespecified dimensions, and delayed tests should assess retention after practice effects have diminished. Report effect sizes and confidence intervals for both successful and unsuccessful probes, rather than emphasizing threshold significance. Where multiple correlated outcomes are studied, provide the multiplicity and dependence structure so readers can assess whether the pattern is selective or consistent.
Evidence synthesis must distinguish original controlled experiments, narrative theoretical reviews, systematic reviews, and meta-analyses. The present monograph is an integrative critical synthesis of identifiable sources, not a registered systematic review with exhaustive search and formal pooled estimates. Reviews like Melby-Lervåg et al. (2016) and Barnett & Ceci (2002) can establish the scope of a debate and study-design moderators, but their conclusions remain tied to the measures and studies included. The next generation of generalization research benefits from adversarial model comparison, cross-laboratory replication, transparent negative results, and explicit measurement of representations and retrieval operations that would distinguish rival accounts.
Integrated Evidence and Open Research Questions
Several findings are jointly informative across the experimental traditions examined here. Responding to untrained stimuli often varies with learned psychological similarity, as formalized by Shepard (1987). Discrimination training can reorganize response gradients, as demonstrated in the tradition of Hanson (1959). Analogical transfer can succeed when structural relations are made accessible, yet source retrieval can be a limiting step, as shown by Gick & Holyoak (1980). Practice organization changes the relationship between immediate proficiency and delayed generalization, while far transfer from narrowly targeted cognitive training often fails rigorous evaluation, as documented in Melby-Lervåg et al. (2016). The shared principle is that successful transfer requires a usable representation, a way to identify its relevance, and compatible conditions for expression.
The major unresolved theoretical questions concern representational structure and causal mechanisms. Under what circumstances does a learner switch from similarity-based interpolation to explicit causal or relational extrapolation? How does training change psychological distance? When do new contexts trigger the retrieval of old knowledge, and when do they trigger inference of a new latent state? Which kinds of variability promote abstraction without overwhelming novice learners? Can measured neural or computational mechanisms predict the degree of transfer across domains and developmental stages? Hybrid theories surveyed by Wu et al. (2025) provide ways to formalize these possibilities, but discriminating experiments must still test predictions beyond the data used to fit each model.
Open methodological questions are equally consequential. Researchers need agreed descriptions of transfer distance, robust cross-task measurement, and methods for distinguishing task-specific improvement from wider change. Educational studies need alignment between tested mechanism and classroom implementation; comparative studies need a principled way of matching task demands across species or systems. Human–artificial investigations must measure supported performance separately from independent retention and transfer. Evidence in one domain cannot simply be declared general across all forms of learning. Progress requires mapping the exact transitions under which knowledge travels, the errors it produces, and the discriminations that make its application adaptive.
Conclusion
Generalization, discrimination, and transfer form a coherent scientific problem: how experience acquired in particular conditions becomes guidance for situations that differ from those conditions. Behavioral experiments quantify response gradients and acquired stimulus control; cognitive theories explain exemplar retrieval, category abstraction, analogy, and causal inference; educational and motor studies test the durability of usable skills across tasks and contexts. The empirical evidence supports substantial, precisely bounded forms of generalization while showing that broad claims of domain-general improvement require much stronger proof than gains on practiced tasks. The scope of transfer is itself the result to be explained.
The most productive synthesis treats similarity, structured knowledge, and contextual retrieval as complementary components of learning. Discrimination identifies where prior knowledge applies and where its application should stop; generalization allows economical use of experience; transfer describes the tested consequences of that use across tasks, settings, and times. Their interaction clarifies why practice can produce both useful abstraction and systematic error. A scientific report is complete when it specifies what was trained, which distinctions were acquired, what changed at test, how success was measured, and which competing mechanisms remain plausible.
Future research can deepen this framework through preregistered transfer gradients, richer stimulus sampling, independent delayed assessments, and experiments contrasting rule use, similarity, and latent-context inference. Studies of learning with artificial systems add new opportunities to separate assistance from competence and to compare mechanisms across human and artificial learners. The enduring conclusion of the research tradition is an operational one: the value of learning is revealed by its demonstrated use under altered circumstances, and the precision of that use depends on the learner's ability to recognize both continuity and meaningful difference.
References
- Barnett, S. M., & Ceci, S. J. (2002). When and where do we apply what we learn? A taxonomy for far transfer. Psychological Bulletin, 128(4), 612–637. https://doi.org/10.1037/0033-2909.128.4.612
- Boldrin, L. S., Quigley, M., & Dymond, S. (2025). Symbolic or derived generalization of fear and avoidance in humans: A systematic review. Journal of Contextual Behavioral Science, 35, 100869. https://doi.org/10.1016/j.jcbs.2024.100869
- Bouton, M. E. (1993). Context, time, and memory retrieval in the interference paradigms of Pavlovian learning. Psychological Bulletin, 114(1), 80–99. https://doi.org/10.1037/0033-2909.114.1.80
- Diamond, A., & Lee, K. (2011). Interventions shown to aid executive function development in children 4 to 12 years old. Science, 333(6045), 959–964. https://doi.org/10.1126/science.1204529
- Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155–170. https://doi.org/10.1207/s15516709cog0702_3
- Gentner, D., Loewenstein, J., & Thompson, L. (2003). Learning and transfer: A general role for analogical encoding. Journal of Educational Psychology, 95(2), 393–408. https://doi.org/10.1037/0022-0663.95.2.393
- 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
- Gick, M. L., & Holyoak, K. J. (1980). Analogical problem solving. Cognitive Psychology, 12(3), 306–355. https://doi.org/10.1016/0010-0285(80)90013-4
- Goldstone, R. L. (1994). Influences of categorization on perceptual discrimination. Journal of Experimental Psychology: General, 123(2), 178–200. https://doi.org/10.1037/0096-3445.123.2.178
- Guttman, N., & Kalish, H. I. (1956). Discriminability and stimulus generalization. Journal of Experimental Psychology, 51(1), 79–88. https://doi.org/10.1037/h0046219
- Hanson, H. M. (1959). Effects of discrimination training on stimulus generalization. Journal of Experimental Psychology, 58(5), 321–334. https://doi.org/10.1037/h0042606
- Harlow, H. F. (1949). The formation of learning sets. Psychological Review, 56(1), 51–65. https://doi.org/10.1037/h0062474
- Holyoak, K. J., & Koh, K. (1987). Surface and structural similarity in analogical transfer. Memory & Cognition, 15(4), 332–340. https://doi.org/10.3758/BF03197035
- Jerrim, J., Lopez-Agudo, L. A., Sims, S., & Marcenaro-Gutierrez, O. D. (2025). Teaching for near transfer: Is maths instruction aimed at schema formation and abstraction associated with pupils' ability to answer unfamiliar maths questions? Learning and Individual Differences, 118, 102609. https://doi.org/10.1016/j.lindif.2024.102609
- Kornell, N., & Bjork, R. A. (2008). Learning concepts and categories: Is spacing the enemy of induction? Psychological Science, 19(6), 585–592. https://doi.org/10.1111/j.1467-9280.2008.02127.x
- Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological Review, 85(3), 207–238. https://doi.org/10.1037/0033-295X.85.3.207
- Melby-Lervåg, M., Redick, T. S., & Hulme, C. (2016). Working memory training does not improve performance on measures of intelligence or other measures of far transfer: Evidence from a meta-analytic review. Perspectives on Psychological Science, 11(4), 512–534. https://doi.org/10.1177/1745691616635612
- Morris, C. D., Bransford, J. D., & Franks, J. J. (1977). Levels of processing versus transfer appropriate processing. Journal of Verbal Learning and Verbal Behavior, 16(5), 519–533. https://doi.org/10.1016/S0022-5371(77)80016-9
- Nokes-Malach, T. J., & Mestre, J. P. (2013). Toward a model of transfer as sense-making. Educational Psychologist, 48(3), 184–207. https://doi.org/10.1080/00461520.2013.807556
- Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115(1), 39–61. https://doi.org/10.1037/0096-3445.115.1.39
- Ritchie, S. J., & Tucker-Drob, E. M. (2018). How much does education improve intelligence? A meta-analysis. Psychological Science, 29(8), 1358–1369. https://doi.org/10.1177/0956797618774253
- Rohrer, D., & Taylor, K. (2007). The shuffling of mathematics problems improves learning. Instructional Science, 35(6), 481–498. https://doi.org/10.1007/s11251-007-9015-8
- Sala, G., & Gobet, F. (2019). Cognitive training does not enhance general cognition. Trends in Cognitive Sciences, 23(1), 9–20. https://doi.org/10.1016/j.tics.2018.10.004
- Schmidt, R. A., & Bjork, R. A. (1992). New conceptualizations of practice: Common principles in three paradigms suggest new concepts for training. Psychological Science, 3(4), 207–218. https://doi.org/10.1111/j.1467-9280.1992.tb00029.x
- Shepard, R. N. (1987). Toward a universal law of generalization for psychological science. Science, 237(4820), 1317–1323. https://doi.org/10.1126/science.3629243
- Simons, D. J., Boot, W. R., Charness, N., Gathercole, S. E., Chabris, C. F., Hambrick, D. Z., & Stine-Morrow, E. A. L. (2016). Do brain-training programs work? Psychological Science in the Public Interest, 17(3), 103–186. https://doi.org/10.1177/1529100616661983
- Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalization, similarity, and Bayesian inference. Behavioral and Brain Sciences, 24(4), 629–640. https://doi.org/10.1017/S0140525X01000061
- Thorndike, E. L., & Woodworth, R. S. (1901). The influence of improvement in one mental function upon the efficiency of other functions (I). Psychological Review, 8(3), 247–261. https://doi.org/10.1037/h0074898
- Wu, C. M., Meder, B., & Schulz, E. (2025). Unifying principles of generalization: Past, present, and future. Annual Review of Psychology, 76, 275–302. https://doi.org/10.1146/annurev-psych-021524-110810