Multiple exemplar training

Exemplars sample meaningful variation

If the goal is identifying fragile packages, training may include boxes, envelopes, bags, different warning labels, sizes, colors, materials, and locations. Relevant features should predict the correct response. Irrelevant features should vary so color or position does not become an accidental shortcut.

Training five nearly identical cards is numerically multiple but may sample little useful variation. Exemplar selection should reflect the range the person will encounter.

Generalization needs a planned target

“Generalization” can mean responding with new materials, people, settings, instructions, or response forms. State which type is expected. A novel item should be identified before the probe and excluded from teaching until the planned test.

If staff prompt or correct the first probe, that stimulus is no longer untrained on later trials. Record the exposure and separate initial probes from subsequent teaching.

Relevant and irrelevant dimensions matter

General case analysis asks which features define the range of real situations. For a community sign, symbol shape and wording may be relevant while font, background color, and sign height vary.

Holth’s review explains strengths and conceptual limits of MET. It cautions that some complex or relation-based performances are not solved simply by adding more direct examples.

MET often appears inside a package

Prompting, modeling, reinforcement, error correction, and feedback may accompany exemplar variation. When outcomes improve, a package study cannot isolate MET as the sole active component.

Marzullo-Kerth and colleagues used multiple material categories with video modeling, prompting, and reinforcement to teach sharing to four autistic children. All showed within-category generalization; one showed across-category generalization. The small treatment-package study supports measuring generalization rather than assuming it.

A fictional exemplar record

Rafael chooses a goal of sorting fragile packages into a padded cart. Training includes eight packages that vary in size, color, material, label location, and sender. The defining features are the approved fragile symbols and handling instructions.

On six predeclared novel-package probes without prompts or feedback, Rafael sorts 5 of 6 correctly. The result is 83.3% for that novel set. It does not establish generalization to every package, workplace, symbol system, or staff instruction. It also cannot show that exemplar variation alone caused the performance.

Probe design protects the claim

For every probe set, record:

• how items differ from training examples
• relevant and irrelevant dimensions
• people, setting, instruction, and response form
• prompts, feedback, and prior exposure
• correct responses divided by eligible probes
• errors by feature and any accessibility barriers

Keep training accuracy, novel probes, maintenance, and daily-use performance separate. An overall percentage can hide success with one category and failure with another.

More variation can also create confusion

Rapid changes in materials or wording may increase effort before the defining relation is clear. Begin with discriminable examples, teach the relevant feature, and add variation according to data. Check whether the person can perceive the features and use the required response.

Provide AAC, interpreters, sensory access, mobility support, breaks, and ordinary help. Eye contact and speech are never prerequisites for generalization when another reliable response works.

Generalization should improve daily fit

Select a skill the person values in settings that matter. Include the people who will support it and ask whether performance feels useful. A correct response in contrived probes may offer little benefit if materials, effort, or consequences differ in daily life.

A recent four-participant study of serial MET for bidirectional naming reported improvement for all participants, mastery for three, and weaker responding during later generalization probes. The pattern again shows why transfer and maintenance need direct measurement.

The BACB BCBA Test Content Outline, 6th edition includes programming for generalization and maintenance in examination content. It does not define an adequate exemplar count or clinical plan.

Decide when to add or stop exemplars

Add examples when errors cluster around a feature missing from training. If every novel flexible pouch is missed, include representative pouches while retaining other variation. Stop adding near-duplicates when probes already show stable performance across the planned range.

Set a review rule in advance. For example, revise the set after two failed novel probes from the same category, then label those items as teaching examples. Preserve the original probe result so improvement is not mistaken for untrained generalization.

Maintenance checks should sample the relevant range after time passes. Everyday performance may also require partner training or environmental changes, especially when a correct response depends on another person recognizing and supporting it.