Generalization and Maintenance: Programming for the Skill to Survive Discharge

Generalization is the occurrence of a learned behavior under conditions other than those in which it was taught — different settings, people, materials, or instructions. Maintenance is its persistence over time after teaching ends. Both are the actual goal of intervention, and both are routinely left to chance.

Stokes and Baer described the prevailing approach in 1977 as train and hope: teach the skill, then hope it shows up elsewhere. Nearly fifty years on it remains the most common practice, and it is why so many programs produce mastery data that families do not recognize at home.

The two kinds of generalization

Stimulus generalization — the behavior occurs in the presence of stimuli other than those used in teaching. A learner who requests a break from the trained RBT also requests one from a substitute, at home, and at school.

Response generalization — the learner emits variations of the trained response that serve the same function. Taught I want a break, the learner also produces can I stop and I need a minute.

Both matter. Programs almost always attend to the first and almost never to the second.

Strategies that actually work

Stokes and Baer's framework remains the practical reference. The strategies that carry most of the weight in clinical practice:

Train sufficient exemplars

Teach with multiple examples rather than one. Multiple instructors, multiple sets of materials, multiple phrasings of the instruction, multiple settings.

This is the highest-yield strategy and the most frequently under-applied. Teaching colour identification with one set of cards produces a learner who identifies those cards. Teaching with six varied sets produces a learner who identifies colour.

How many exemplars is enough is an empirical question per learner — teach, probe an untrained exemplar, and add exemplars until untrained ones are correct.

Program common stimuli

Deliberately include features of the natural environment in teaching. If the goal is requesting at school, use school materials, school furniture, school routines and school vocabulary during teaching.

Train loosely

Vary the non-essential aspects of teaching on purpose — tone of voice, instruction wording, position in the room, time of day, which hand materials are presented with. Rigid consistency in teaching produces rigid stimulus control, which is exactly what you do not want at the end.

This runs against the instinct toward procedural consistency, and the resolution is sequencing: tight during acquisition, deliberately loose once the response is established.

Introduce to natural maintaining contingencies

The most durable strategy. Teach responses that the natural environment will reinforce without you.

A learner taught to request help will keep requesting help because people respond to it. A learner taught to complete a worksheet for an arbitrary token will stop when the tokens stop. Where possible, choose target behaviors the ordinary world already reinforces — and where the natural contingency is weak, teach the people in that environment to deliver it.

Use indiscriminable contingencies

Thin reinforcement to intermittent and unpredictable schedules, and delay reinforcement, so the learner cannot discriminate which responses will be reinforced. Behavior maintained on a lean, unpredictable schedule is far more resistant to extinction than behavior on a dense predictable one.

Mediate generalization

Teach the learner a portable mediator that travels between settings — a self-management system, a written or visual reminder, a self-instruction script. The mediator is present everywhere the learner is.

Train to generalize

Treat generalization itself as the response class to reinforce. Explicitly reinforce novel applications and variations of the skill rather than only trained forms.

Maintenance

Maintenance is generalization across time, and it needs its own programming:

• Thin reinforcement systematically before ending teaching. A skill maintained on continuous reinforcement will not survive the natural environment's much leaner schedule.
• Overlearn. Continue practice past the initial mastery criterion. Skills taught to bare criterion decay fastest.
• Build fluency, not just accuracy. Fluent responding is more durable than accurate but slow responding, and it is more likely to compete successfully with other behavior in the natural environment.
• Schedule maintenance probes at expanding intervals after mastery — one week, one month, three months. Without scheduled probes, skill loss is discovered accidentally, usually months late.

Measuring it

Generalization that has not been measured has not been demonstrated. Two mechanisms:

Generalization probes — conduct the target in an untrained condition, with no prompting and no reinforcement beyond what is naturally available. Run them at baseline, periodically during teaching, and at mastery. Score them separately from teaching data so untrained performance never gets mixed into the trained mastery figure.

Maintenance probes — re-run mastered targets at expanding intervals to confirm the skill is still there.

Both need to be scheduled rather than remembered. In practice this is where most programs lose the thread: mastery is recorded, the target closes, and nothing ever checks it again. A program register that surfaces mastered targets for probing at one week, one month and three months turns maintenance from an intention into a routine — and mastered targets that quietly decayed are one of the most common findings when anyone actually goes back and checks.

Writing it into the program

A skill acquisition program that addresses generalization states, before teaching begins:

1. Which dimensions will be varied during teaching — people, settings, materials, instructions.
2. The generalization criterion, stated as performance in untrained conditions rather than trained ones.
3. The probe schedule, both generalization and maintenance.
4. Who in the natural environment will be taught to support the behavior, and how.

Programs that specify a mastery criterion but no generalization criterion are declaring victory at the point where the real work starts.

The summary

Generalization is programmed, not hoped for. Train with multiple exemplars, include natural-environment stimuli, loosen deliberately once the response is established, and prefer behaviors the ordinary world will reinforce on its own. Thin and delay reinforcement to make contingencies indiscriminable, give the learner a portable mediator, and reinforce novel variations. For maintenance, overlearn, build fluency, thin before you exit, and probe on a schedule. Define the generalization criterion in untrained conditions — because mastery measured only where you taught is not mastery.