Response Generalization

Abstract

Response generalization is a form of psychological generalization in which reinforcing one response increases the probability of other, untrained responses that resemble it or serve the same function. It is the response-side complement of stimulus generalization: where stimulus generalization concerns one response spreading across similar stimuli, response generalization concerns one reinforcement spreading across similar responses. The phenomenon follows directly from the operant, which Skinner defined not as a single movement but as a class of responses grouped by their common effect on the environment. A distinct line of work established that variability itself can be reinforced as an operant dimension, so that an organism can be trained to emit novel and unpredictable response forms. This article develops the response class, response induction, variability as a generalized operant, the applied technology for programming generalization, and the theories that tie them together.

Keywords: response generalization, response class, response induction, operant variability, programming generalization

Response generalization is the tendency for the reinforcement of one response to strengthen other responses that share its form or its function, even when those other responses were never themselves reinforced (Skinner, 1938). It is the mirror image of stimulus generalization on the response axis: stimulus generalization asks how a single response transfers across similar stimuli, while response generalization asks how a single reinforcement transfers across similar responses to the same stimulus. The two are measured on complementary scales and together define the surface over which learning spreads. The phenomenon is adaptive for the same reason its stimulus counterpart is. No response recurs in exactly its original form, because muscles fatigue, positions shift, and no two movements are ever identical; an organism that could repeat only the precise movement it was first reinforced for would rarely be reinforced again. Response generalization is what lets a reinforced act be replaced by its many variants, and it is the behavioral raw material from which an organism assembles flexible, novel, and adaptive action. Its systematic study began with the recognition that the unit strengthened by reinforcement is never a single movement but a class of movements, and it now spans the analysis of response variability as a reinforceable dimension and an applied technology for making trained behavior generalize to untrained forms and settings (Stokes & Baer, 1977). The sections below develop the concept and its measurement, its relation to stimulus generalization, the response class that underlies it, variability as a generalized operant, the programming of generalization in applied settings, and the theories that unify these strands.

What Response Generalization Is

Response generalization is defined by a relation between response similarity and the spread of reinforcement: strengthening one response raises the strength of other responses in proportion to how much they resemble the reinforced one in form or in function. The classical term for the same fact is response induction, the observation that reinforcing one member of a set of related responses induces the others. The phenomenon is visible whenever a single reinforced act is measured not as present-or-absent but along a graded response dimension. When a rat is reinforced for pressing a lever, it does not press in one invariant way; the presses vary in force, in duration, in the position on the bar that is struck, and in the paw used, and the frequencies of these variants trace an orderly distribution peaked at the values that have been reinforced (Antonitis, 1951). That distribution is a response-generalization gradient, and it is the response-side analogue of the stimulus-generalization gradient that Guttman and Kalish plotted across a stimulus dimension (Guttman & Kalish, 1956). What makes response generalization a single phenomenon rather than a scattering of separate observations is this orderliness: the untrained variants are not emitted at random but in a graded relation to the reinforced value, strongest near it and weaker farther away. The phenomenon also gives operant learning its reach. Because the motor world never presents the identical opportunity twice, an organism confined to the exact reinforced movement would have learned nothing usable; response generalization is the bridge from a particular reinforced act to the class of acts that accomplish the same end, and its breadth determines how flexibly a learned function can be expressed.

Response Generalization Versus Stimulus Generalization

Response generalization and stimulus generalization are complementary halves of the same three-term contingency, and keeping them distinct is the first step to understanding either. A discriminated operant is a relation among an antecedent stimulus, a response, and a consequence; generalization can occur on either the stimulus side or the response side of that relation, and the two spreads are logically independent. Stimulus generalization holds the response constant and varies the stimulus: a response reinforced in the presence of one discriminative stimulus is emitted to other, similar stimuli, and the frequency of that one response declines with distance from the trained stimulus along a stimulus dimension (Guttman & Kalish, 1956). Response generalization holds the stimulus constant and varies the response: in the presence of the same stimulus, reinforcing one response raises the frequency of other, similar responses, and those untrained responses decline in frequency with distance from the reinforced form along a response dimension (Antonitis, 1951). The stimulus side was described first, in Pavlov's account of how a conditioned reflex established to one stimulus irradiates to similar stimuli, and the response side was recognized as its formal complement once the operant was understood as a class (Pavlov, 1927). The symmetry is exact but the two are not interchangeable, because a procedure can widen one while narrowing the other. Differential reinforcement of a specific response topography sharpens the response gradient without necessarily affecting the stimulus gradient, and discrimination training between stimuli sharpens the stimulus gradient without touching the response gradient. This independence is why the applied literature treats stimulus generalization and response generalization as separate targets that must each be programmed, and why a treatment that transfers across settings, a stimulus achievement, need not transfer across response forms, a response achievement (Stokes & Baer, 1977). Confusing the two leads to the mistaken expectation that arranging for one automatically secures the other.

The Response Class

Response generalization is intelligible only because the operant is a class, and the response class is the concept that carries the entire phenomenon. Skinner drew the fundamental distinction between a response as a single instance of behavior, a particular movement occurring once, and the operant as a class of such instances grouped by a common relation to the environment (Skinner, 1938). What defines membership in the class is not the topography, the physical form of the movement, but the function, the effect the movement has, typically the production of a reinforcer. A rat's lever press is one operant however it is executed, with the left paw or the right, from above or from the side, hard or soft, because all these topographies close the same switch and produce the same food. This is the functional response class: a set of responses that are functionally equivalent because they operate on the environment in the same way and are maintained by the same consequence. Reinforcement strengthens the class, not the instance, and this is precisely why reinforcing one member raises the probability of the others. Response generalization is therefore not a separate mechanism added onto operant conditioning but a direct consequence of what the operant is: to reinforce a class is to strengthen every member, including members that have not yet occurred. Skinner made the same point about the flexibility of behavior, noting that the organism's tendency to vary the form of a response while preserving its function is what allows behavior to remain effective as circumstances change (Skinner, 1953). The class can also be defined more abstractly than by immediate physical effect. A generalized operant is a functional class whose members share not a common movement but a common relational property, such as imitating any modeled action or matching any sample; such classes are the higher-order form of the same idea and are central to modern accounts of complex behavior. The demonstration below plots a response-generalization gradient over a response dimension and lets the reader vary which response form is reinforced and how sharply the class is differentiated.

Variability as a Generalized Operant

If reinforcement strengthens a response class, a striking question follows: can variability itself, the tendency to emit different forms rather than the same one, be a reinforceable property of behavior? The classical view held that variability is merely what remains when reinforcement is withdrawn, the undifferentiated scatter of responding that extinction unmasks, and Antonitis had shown that reinforcement narrows response variability while extinction widens it (Antonitis, 1951). Page and Neuringer overturned the assumption that variability is only a residue by making it the explicit criterion for reinforcement (Page & Neuringer, 1985). They reinforced pigeons for producing sequences of eight pecks across two keys only when the current sequence differed from recent ones, so that reinforcement was contingent on variability itself rather than on any particular sequence. The birds learned to respond highly variably, generating sequences that approached the statistical unpredictability of a random generator, and crucially their variability rose and fell with the contingency: when variation was required they varied, and when a specific sequence was reinforced they repeated it. Variability behaved, in short, like any other operant dimension, controlled by its consequences. Neuringer's later synthesis assembled the evidence that operant variability is a robust, general, and functionally important dimension of behavior, reinforceable across species and response types, sensitive to schedules, and implicated in problem solving, creativity, and the deficits of clinical conditions (Neuringer, 2002). Reinforced variability is a generalized operant in the strong sense: the reinforced property is not a fixed form but the abstract relation of being-different-from-previous-responses, a property that can only be satisfied by continually generating new members of the class. The claim has not gone unchallenged. Nergaard and Holth reviewed the supporting evidence critically and argued that much of what is attributed to a variability operant can be explained by the differential reinforcement of specific, low-frequency response patterns and by stochastic emission, so that whether variability is truly a directly reinforceable dimension, rather than an artifact of how the contingency is arranged, remains an open theoretical question (Nergaard & Holth, 2020). The demonstration below reinforces variability under an adjustable threshold and shows the distribution of emitted response sequences broaden or collapse as the contingency changes.

Programming Response Generalization

For applied behavior analysis the central practical fact about generalization is that it usually does not happen by itself. A skill taught in a clinic with one therapist and one set of materials tends to stay tied to that setting and that response form unless something is done to spread it, and the field's foundational insight was that generalization must be actively programmed rather than passively hoped for. Stokes and Baer surveyed the applied literature and named the prevailing practice an implicit technology of generalization, a set of tacit procedures that worked when they worked but were rarely planned, and they organized them into an explicit set of tactics (Stokes & Baer, 1977). Among these, two bear directly on response generalization. Training sufficient exemplars teaches not one response form but several members of the target class until the untrained members appear, on the logic that reinforcing a few instances of a functional class recruits the whole class. Training loosely, deliberately varying the antecedents and the accepted response forms during teaching rather than reinforcing one rigid topography, builds a broad response class from the outset and prevents the narrowing that tight differential reinforcement produces. The same logic connects to reinforced variability: explicitly reinforcing varied responding is a way to build a response class wide enough to include novel forms, which is valuable when the therapeutic goal is flexible or creative behavior rather than a single fixed skill. The technology has continued to develop and to reach new delivery formats. Shawler and colleagues, for instance, used telehealth to teach caregivers behavior-analytic skills and programmed for the generalization of those skills to untrained situations, showing that the classic tactics transfer to remote instruction and to the training of the trainers themselves (Shawler et al., 2023). The demonstration below contrasts training a single exemplar with training several, and shows how the number of trained members of a response class governs how much of the untrained class comes along.

Theories of Response Generalization

The theoretical accounts of response generalization form a progression from the concrete functional class to increasingly abstract classes defined by relations. The foundational account is Skinner's operant itself: because the operant is a class defined by function, reinforcement of any member strengthens the class, and response generalization is the automatic expression of that fact (Skinner, 1938; Skinner, 1953). On this view no additional theory is needed for the basic phenomenon; the gradient of untrained response forms is simply the shape of the class as reinforcement has carved it. The variability research extends the account by showing that the reinforceable property of a class need not be a topography at all but can be an abstract statistical relation among successive responses, which makes the generalized operant a genuine theoretical category rather than a description of a movement (Page & Neuringer, 1985; Neuringer, 2002). The most abstract extension treats relations themselves as the reinforced units. Dixon and colleagues, working within relational frame theory, provided evidence from children with autism that derived relational responding, the ability to respond to stimuli in terms of relations such as sameness or opposition that were never directly trained, behaves as a generalized operant: it can be established through reinforcement of a subset of relations and then generalizes to untrained relations, exactly as a response class should (Dixon et al., 2021). This positions the most complex human abilities, including aspects of language and reasoning, as high-order response classes governed by the same generalization logic that describes a rat's lever press. A separate theoretical connection runs to the formal treatment of generalization in general. Shepard's universal law, that generalization decays exponentially with distance in an internal psychological space, was framed for stimulus generalization but its logic, an organism inferring which variants share a consequence, applies equally to inferring which response forms share a reinforcer, suggesting that response and stimulus generalization may be two applications of a single principle of inference under uncertainty (Shepard, 1987). What unites these accounts is the response class: whether the class is defined by a shared movement, a shared statistical property, or a shared relation, reinforcement acts on the class and generalization is the spread of that action to its untrained members.

Worked Example

The claim that variability is a measurable, reinforceable dimension rests on being able to quantify how variable a set of responses is, and working the measure by hand shows what a variability contingency actually rewards. Page and Neuringer used an uncertainty statistic, U, that ranges from 0 when responding is completely stereotyped to 1 when every possible response form is equally likely. For a set of k possible response forms emitted with relative frequencies p, the statistic is U = -Σ p log2(p) / log2(k). Consider an organism with four available response forms. Under a contingency that reinforces one particular form, suppose the emitted distribution is heavily concentrated, p = [0.70, 0.10, 0.10, 0.10]. The numerator is -(0.70 · log2 0.70 + 3 · 0.10 · log2 0.10) = -(0.70 · -0.5146 + 0.30 · -3.3219) = -(-0.360 + -0.997) = 1.357, and dividing by log2 4 = 2 gives U = 0.68. Now suppose a variability contingency instead reinforces differing from recent responses, and the emitted distribution becomes uniform, p = [0.25, 0.25, 0.25, 0.25]. Each term is 0.25 · log2 0.25 = 0.25 · -2 = -0.50, the numerator is -(4 · -0.50) = 2.0, and dividing by log2 4 = 2 gives U = 1.0. The variability contingency has driven the uncertainty statistic from 0.68 to its ceiling of 1.0, and it did so without reinforcing any single form more than another; what was reinforced was the spread across the class itself. The lesson is that response generalization can be measured on the same footing as its stimulus counterpart, and that variability is not the mere absence of control but a dimension that a contingency can push up or down, exactly as Page and Neuringer demonstrated (Page & Neuringer, 1985).

Discussion

Response generalization has traced a path from a simple corollary of the operant to a full theory of how flexible behavior is built, and the through-line is the response class. The recognition that the operant is a functional class, not a fixed movement, made response generalization automatic: to reinforce a class is to strengthen its untrained members, and the graded distribution of response forms that Antonitis measured is nothing more than the shape of that class (Skinner, 1938; Antonitis, 1951). Holding the concept alongside stimulus generalization clarified that the two are independent spreads on opposite sides of the same contingency, each requiring its own analysis and its own programming (Guttman & Kalish, 1956). The variability research then showed that the reinforceable property of a class can be an abstract statistical relation rather than a topography, turning the generalized operant into a substantive theoretical category and connecting response generalization to problem solving and creativity (Page & Neuringer, 1985; Neuringer, 2002). The applied tradition supplied the practical counterpart, an explicit technology for making trained behavior generalize across forms and settings rather than assuming it would (Stokes & Baer, 1977). At the theoretical frontier, relational frame theory has extended the same logic to derived relational responding, positioning language and reasoning as high-order response classes governed by generalization (Dixon et al., 2021). Yet the account is not closed. Whether reinforced variability is a directly controllable operant dimension or an artifact of how variability contingencies are arranged remains genuinely disputed, and the answer bears on how far the generalized-operant idea can be pushed (Nergaard & Holth, 2020). What began as the observation that a reinforced act comes in many forms is now a framework in which the unit of learning is a class and the reach of learning is the spread of reinforcement across it.

Current Directions

Contemporary work on response generalization concentrates where the basic concept meets clinical and translational application. The variability program has moved from establishing that operant variability exists to interrogating its foundations, and a live methodological debate asks whether the classic demonstrations isolate a true variability dimension or whether differential reinforcement of specific low-probability patterns and stochastic emission can account for the same results, a question with direct consequences for interventions that aim to increase behavioral flexibility (Nergaard & Holth, 2020). In applied behavior analysis the generalized-operant framework has become central to the treatment of autism, where derived relational responding is treated as a trainable higher-order response class and used to build language repertoires that extend beyond directly taught relations (Dixon et al., 2021). A parallel practical strand is the migration of the Stokes and Baer technology to new delivery formats: programming generalization through telehealth, and generalizing the behavior of caregivers and other mediators rather than only the target client, extends the reach of intervention and tests whether the classic tactics survive remote and second-order application (Shawler et al., 2023). Threading through these programs is the older theoretical question of whether response generalization and stimulus generalization are two faces of a single inferential principle, of the kind Shepard proposed for the stimulus side, and whether the exponential form he derived has a response-side analogue (Shepard, 1987). The enduring aim is an account in which the response gradient measured in an operant chamber, the reinforced variability seen in a sequence task, the derived relations trained in a language program, and the generalization engineered in a clinic are all expressions of the same class-based process.

Common Misconceptions

Response generalization is the same thing as stimulus generalization.

They are complementary but independent. Stimulus generalization holds the response constant and spreads it across similar stimuli; response generalization holds the stimulus constant and spreads reinforcement across similar responses. A procedure can widen one while leaving the other untouched, which is why applied work programs each separately and why transfer across settings does not guarantee transfer across response forms (Stokes & Baer, 1977).

Variability is simply the absence of reinforcement or control.

Extinction does increase variability, but that is not the whole story. When variability itself is made the criterion for reinforcement, organisms come to respond more variably, and their variability rises and falls with the contingency, behaving like any other operant dimension. Variability can therefore be a product of reinforcement, not only of its withdrawal (Page & Neuringer, 1985).

Generalization happens on its own once a skill is learned.

The founding lesson of the applied literature is the opposite: generalization is usually limited and must be deliberately programmed. Training a single exemplar in a single setting tends to produce behavior bound to that exemplar and setting, and tactics such as training sufficient exemplars and training loosely are needed to make the untrained members of a response class and untrained settings come along (Stokes & Baer, 1977).

Glossary

Behavioral variability.

The extent to which successive responses differ from one another in form or sequence; shown to be a dimension that reinforcement can increase or decrease rather than merely a residue of extinction.

Derived relational responding.

Responding to stimuli in terms of relations such as sameness or opposition that were never directly trained; in relational frame theory it is treated as a generalized operant central to language.

Functional response class.

A set of responses that are grouped together because they produce the same effect on the environment and are maintained by the same consequence, regardless of differences in their physical form.

Generalized operant.

A response class whose members share not a common movement but a common abstract property or relation, such as varying from previous responses or imitating any modeled action.

Lag schedule.

A reinforcement schedule that delivers reinforcement only when the current response differs from a specified number of preceding responses; the standard arrangement for reinforcing variability.

Operant.

A class of responses defined by their common effect on the environment and strengthened as a class by reinforcement, as distinct from a single instance of behavior.

Programming generalization.

The deliberate use of teaching tactics to make trained behavior spread to untrained response forms and settings, rather than expecting such transfer to occur spontaneously.

Response class.

A group of responses treated as a unit because they share a defining property; the entity on which reinforcement acts and across which response generalization spreads.

Response differentiation.

The narrowing of a response class by reinforcing only a restricted range of response forms, which sharpens the response-generalization gradient around the reinforced value.

Response generalization.

The spread of reinforcement from one response to other untrained responses that resemble it in form or function; the response-side complement of stimulus generalization.

Response induction.

The classical term for response generalization: reinforcing one member of a set of related responses induces the emission of the others.

Stimulus generalization.

The transfer of a single conditioned response to stimuli resembling the training stimulus; the stimulus-side complement of response generalization.

Sufficient exemplars.

A generalization tactic that trains several members of a target response class until the untrained members appear, exploiting the class structure of the operant.

Topography.

The physical form of a response, its movement and shape; members of a functional response class can differ in topography while sharing a function.

U-value.

An uncertainty statistic ranging from 0 for completely stereotyped responding to 1 for maximally variable responding, used to quantify behavioral variability.

Key Researchers

Donald M. Baer (1931-2002). Professor at the University of Kansas; with Stokes he wrote the founding statement of a technology of generalization, converting a set of tacit practices into an explicit set of programming tactics. Wikipedia - Memorial - ABA Biography

Per Holth. Professor Emeritus at OsloMet - Oslo Metropolitan University; with Nergaard he produced the critical review questioning whether behavioral variability is truly a directly reinforceable operant dimension. Faculty Page - Google Scholar

Allen Neuringer. Professor Emeritus at Reed College; he established that behavioral variability is itself a reinforceable operant dimension and synthesized the evidence, functions, and theory of operant variability. Faculty Page - Wikipedia - CV

B. F. Skinner (1904-1990). Professor at Harvard University; he defined the operant as a functional class of responses, the concept from which response generalization follows directly. Wikipedia - Wikidata - Britannica

Trevor F. Stokes. Professor at James Madison University; first author of the foundational analysis of programming generalization and a career-long contributor to the generalization of behavior change. Faculty Page - Google Scholar

Frequently Asked Questions

What is response generalization?
Response generalization is the tendency for the reinforcement of one response to strengthen other, untrained responses that resemble it in form or function, so that a single reinforced act spreads to the class of acts that accomplish the same end (Skinner, 1938).

How does response generalization differ from stimulus generalization?
They are complementary sides of the same contingency: stimulus generalization holds the response constant and spreads it across similar stimuli, whereas response generalization holds the stimulus constant and spreads reinforcement across similar responses. The two are independent and each must be analyzed and programmed separately (Guttman & Kalish, 1956; Stokes & Baer, 1977).

What is a response class?
A response class is a group of responses treated as a single unit because they share a defining property, most often a common effect on the environment. Reinforcement strengthens the class rather than the individual instance, which is why reinforcing one member raises the probability of the others (Skinner, 1938).

What is response induction?
Response induction is the classical term for response generalization, naming the observation that reinforcing one member of a set of related responses induces the emission of the other members of the set (Antonitis, 1951).

Can variability itself be reinforced?
Yes. When reinforcement is made contingent on a response differing from recent responses, organisms come to respond highly variably, and their variability rises and falls with the contingency, showing that variability behaves as a reinforceable operant dimension (Page & Neuringer, 1985; Neuringer, 2002).

Is reinforced variability a settled fact?
No. A critical review has argued that many demonstrations of a variability operant can be explained by the differential reinforcement of specific low-probability patterns and by stochastic emission, so whether variability is a directly reinforceable dimension remains an open theoretical question (Nergaard & Holth, 2020).

How is response generalization programmed in applied settings?
Because generalization is usually limited, it must be programmed deliberately, using tactics such as training sufficient exemplars, teaching several members of a target response class until the untrained members appear, and training loosely, varying the antecedents and accepted response forms to build a broad class from the start (Stokes & Baer, 1977).

How does response generalization relate to language and reasoning?
Relational frame theory treats derived relational responding, responding to untrained relations such as sameness or opposition, as a generalized operant established through reinforcement of a subset of relations, positioning complex verbal abilities as high-order response classes governed by generalization (Dixon et al., 2021).

References