Generalization in ABA: Stimulus vs Response, and How to Program It

A skill that only works in the room where it was taught is not a skill, it is a party trick. Generalization is the difference, and the uncomfortable finding behind fifty years of work on it is that it does not reliably happen on its own. You get it because someone planned for it.

Two kinds, and the exam cares which

Stimulus generalization is the same response under new conditions. A learner taught to greet their therapist greets a teacher, a grandparent, and the neighbor. The behavior is unchanged; the antecedent conditions are new.

Response generalization is new forms of the response under the trained condition. A learner taught to say "I want juice" also produces "juice please," "can I have juice," and a point at the fridge. The condition is unchanged; the behavior takes forms nobody drilled.

The reliable way to tell them apart under time pressure is to ask what moved. If the room, the person or the materials changed, it is stimulus. If the words or the movements changed, it is response.

Maintenance is a third thing, not a kind of generalization

Maintenance is the behavior still occurring after the intervention has stopped. It runs along time rather than across conditions, which is why the Test Content Outline gives it its own task. A skill can generalize beautifully across five settings in March and be gone by June. Those are separate questions and they need separate probes.

Train and hope is not a plan

In 1977 Stokes and Baer reviewed the generalization literature and sorted every study into nine categories. The first is called Train and Hope, and the name is the argument: teach, then check whether generalization happened, and treat it as a pleasant surprise when it did. Their point was that generalization had been treated as what you get when discrimination training is sloppy, rather than as something with a technology of its own.

The remaining eight are that technology:

• Sequential modification. Find where it did not generalize, then go and train there too.
• Introduce to natural maintaining contingencies. Teach behaviors the everyday environment will already reinforce, and stop propping them up artificially.
• Train sufficient exemplars. Teach across several people, settings or materials rather than one, until the learner responds to untrained ones.
• Train loosely. Deliberately vary the irrelevant parts of teaching so no single incidental feature comes to control the response.
• Use indiscriminable contingencies. Make it unclear when reinforcement is coming, usually by thinning to an intermittent schedule.
• Program common stimuli. Move a salient feature of the teaching setting into the natural one, or the reverse.
• Mediate generalization. Give the learner something portable that travels with them, such as a self-instruction or a self-management system.
• Train "to generalize." Treat generalizing itself as the response class, and reinforce novel application directly.

Train sufficient exemplars and train loosely carry most of the practical weight, and they cost nothing but planning.

What loose training looks like in a session

Target: the learner labels a cup. Tight training uses the same blue plastic cup, the same table, the same phrasing, and the same therapist. It produces a learner who labels that cup.

• Rotate the exemplar: paper cup, mug, glass, travel cup, a cup in a picture book.
• Rotate the instruction: "what is it," "tell me what you see," "what's this one," and sometimes nothing but an expectant look.
• Rotate who asks, and where. Two staff, a parent, the kitchen, the car.
• Keep the target constant and let everything else drift.

That last line is the whole method. Anything held rigidly constant during teaching is a candidate for controlling the response, and you will not find out which one did until the learner meets a cup you did not use.

Where this sits on the exam

Task B.14 of the 6th Edition Test Content Outline asks you to distinguish stimulus from response generalization, and B.15 covers response maintenance separately. Both sit in Concepts and Principles, which is the largest task domain on the exam at 24 tasks. Scenario items here are usually solvable by the single question above: did the conditions change, or did the behavior?