Skip to content

Teach response precision and new associations with minimal errors, using mediators and stimulus fading

Principle · All principles
Evidence · 15 claims (7 for, 8 mixed) · 22 studies (7 review, 6 causal, 5 theoretical, 4 quant-synthesis), q1–q4 · 5 of 22 report an effect size · 9 claims rest on one study

Conditional relationship

This page owns one decision: whether to prevent errors while a learner acquires a precise response or a new association, rather than letting errors happen and correcting them. Two kinds of target fall under it. Response precision is a performance that has to come within a tolerance (letter size and spacing in handwriting, timing in music, a measurement or a dose at work): a procedure or skill-performance judged against a criterion. New associations are pairings to be recalled or discriminated (a word and its meaning, a symbol and its value, a colour and its digit in the resistor code): verbal-association in this wiki's terms.

The proposed relationship is conditional. For a learner who is novice on this specific target, instruction that keeps responses almost always correct, by (a) setting the first performance criterion wide enough to include most of what the learner already does and then narrowing it in small steps, (b) linking each new item to something the learner already knows through a mediator (a keyword, an acrostic, an existing code), and (c) giving cues that make the correct response near-certain and then removing them across trials (stimulus fading), is expected to produce accurate unaided performance (immediate-performance, skill-performance) with fewer of the costs that frequent errors carry: rehearsing the wrong response, avoidance, and loss of motivation when the criterion tightens abruptly. Errors that do occur are corrected at once with the correct response. Whether the result holds at delayed-retention depends on how fast the cues come away, and there the evidence below pulls against keeping support high.

It is conditional on four things: the kind of target (the argument is made for precision and associations only; for a concept, principle or a higher-order strategy, attempting first and erring can serve conceptual learning better); the learner's starting response on this task (support that helps a novice can be redundant for an advanced learner); the horizon (cue support that keeps practice error-free can weaken retention a day or more later); and who corrects an error, and how soon.

No claim in the wiki tests this relationship with human learners. The page's source is a 1966 conceptual analysis (Glaser) whose evidence for errorless training is Terrace's work with pigeons, reported second-hand. What the wiki does hold are neighbouring tests: mnemonic instruction against other teaching for students with mild disabilities (recall of facts and vocabulary), corrective feedback after retrieval against none, explicit instruction against unassisted discovery, and a retrieval-practice literature in which less cue support during practice gives better retention. The tolerance-contraction half (shaping precision) has no claim behind it here at all.

How it sits beside converted pages: Guided Practice owns the practice phase after a model for structured tasks generally; this page is the error-minimizing end of it for two narrow kinds of target. Error Analysis owns using an error as an object of study, and Immediate Feedback owns when correction arrives. Productive Failure owns the opposite decision for new concepts, attempting before instruction. Scaffolding and Fading owns how support is raised and withdrawn contingently; stimulus fading here is one form of it. Deliberate Practice owns refining a performance the learner can already carry out roughly.

Default design, while the relationship is untested

The earlier page's guidance, made concrete. Every step is a design proposal; its evidence status is beside it. Quantities are untested proposals unless a claim is cited for them.

For response precision (shaping a tolerance)

  1. Measure current variation before setting a criterion. Take 5–10 unaided attempts and score each against the final criterion (letter height in millimetres, beat timing, measured value). Record the spread. Untested proposal.
  2. Set the first criterion to include most current attempts. The earlier page: the starting range "must include much of the learner's existing variation". Proposal: set it so that about four of the last five attempts already meet it. Untested; from Glaser's argument (The usefulness of induction and errorful learning varies with the type of terminal task being taught [+W], a conceptual analysis).
  3. Narrow the criterion in small steps, never abruptly. Proposal: narrow by about a fifth of the remaining gap to the final tolerance when 8 of the last 10 attempts meet the current one; widen back one step if fewer than 6 of 10 do. The reason given by the earlier page is that an abrupt tightening risks extinction of the response and loss of motivation; that is Glaser's argument, untested here.
  4. Show the learner the current band and their attempts against it (a template line on the page, a metronome click, a gauge), so each attempt is judged at once. Untested proposal.

For new associations (mediators and fading)

  1. Give each new item a mediator the learner already knows: a keyword that sounds like the new word and an image joining it to the meaning, an acrostic, an existing code (the earlier page's resistor colour code). Present the mediator ready-made for young learners and learners who struggle to remember; let confident learners make their own. Evidence: Mnemonic Instruction Improves Recall For Students With Disabilities [+S] (students with mild disabilities, recall of facts and word meanings; the claim page's comparison is general instruction, and its outcomes are recall, not understanding).
  2. Start with a cue that makes the correct answer near-certain, then remove it in steps. Show the mediator image and the first letters of the answer, then fewer letters, then the mediator alone, then nothing. Evidence for fading the cue rather than keeping it: Initial retrieval conditions that provide less cue support, such as free recall rather than recognition or fewer letter cues, tend to produce better retention despite lower initial success [~M] (a review chapter reporting word-list and foreign-word-pair experiments: a diminishing-cues condition beat study-only and accumulating cues). The same claim limits the design: fewer cues produced better retention despite more failures, so do not hold cues at full strength to keep practice error-free. Errorless discrimination as a technique: Errorless discrimination training avoids emotional responses and yields more effective transfer than errorful training (Terrace's pigeon studies) [+W] (pigeons, reported second-hand; carried to people untested).
  3. Correct every error at once with the correct response, then bring the item back. Say or show the right answer, have the learner produce it, and re-present the item two or three items later. Evidence: Feedback Enhances Retrieval Practice [+S] (classroom quizzing with corrective feedback against without; a word-pair experiment in which the correct answer after a wrong response raised retention a week later, while feedback after correct responses made little difference). For young children, make the correction overt (show and say the correct answer, not only "wrong"): Overt correction of errors speeds young children's learning compared with knowledge of correctness alone [+W] (reported second-hand; adults did not learn faster with overt correction).
  4. Never leave a wrong choice uncorrected in multiple-choice practice. Evidence: Taking initial multiple-choice tests without feedback can lead students to later produce the incorrect lure answers they selected [+W] (college students, recall tests, reported in a review chapter).

For both

  1. Keep the approach to precision and associations. For a concept, rule or strategy, use the Productive Failure or Guided Practice designs instead. Evidence: Productive Failure Improves Conceptual Learning [~S] (attempting first, with errors, beat instruction first on conceptual knowledge and transfer, and made no difference on procedural knowledge); Self-discovery through errorful exploration is basically incompatible with error minimization in instruction [~W] (Glaser's argument that the two cannot be combined in one sequence).
  2. Guide novices; do not leave the first attempts to unaided trial and error. Evidence, carried from conceptual and procedural learning: Minimal guidance is less effective for novices than explicit instruction [+M] (164 studies pooled: explicit instruction beat unassisted discovery; discovery with feedback, worked examples or scaffolding did well).

Dose and progression across sessions (untested proposals):

  • Session length. Association practice in blocks of 10–15 minutes, 8–12 new items per block; precision practice in blocks of 5–10 minutes of attempts, inside a longer lesson.
  • Early sessions (first one or two): mediator shown with every new item, full cue, correction at once; precision band wide (step 2).
  • Middle sessions: open each session with the previous session's items at the cue level one step weaker than where they ended; the band starts one step wider than where the last session finished, then narrows within the session.
  • Late sessions: no cue and no mediator shown; old and new items mixed; the band at or near the final tolerance; the learner says or makes the mediator only if they need it.
  • Moving on. An association item retires from daily practice when it is recalled unaided, first time, at the start of two separate sessions at least a day apart; a precision target is met when 8 of 10 attempts fall within the final tolerance at the start of a session, before any warm-up. Then space the reviews out (see Spaced Learning).

When a condition is not met: if no one can see each response (a large class, self-paced work), use items whose correctness software or an answer key can show immediately, with the correct answer displayed after every miss; if the learner already succeeds unaided on the first items, skip the cues and the wide band and go straight to mixed, unaided practice.

Fitting the design to a situation

The three facts that most change the decision:

  • What kind of target it is. A precise response or an arbitrary pairing suits error minimization; a concept or rule the learner must understand may be better served by an attempt that fails first. If unstated, ask: is there a single correct response (or a tolerance) for each item, or must the learner explain or apply an idea to new cases?
  • The learner's starting response on these items. If unstated, ask: before any teaching, how many of ten items (or how many of ten attempts within the final tolerance) does the learner get right unaided?
  • The horizon and the test conditions. Error-free practice with full cues can look good at the end of a session and fade by the next week. If unstated, ask: when must the learner produce the response, with what aids, and how will that be checked?
If the brief says… Then change… Basis
Target is a concept, principle or transfer to new problems, not a pairing or a tolerance Do not use this design; let learners attempt problems first and teach from their attempts, or use guided practice with explanation prompts Productive Failure Improves Conceptual Learning [~S] (conceptual and transfer outcomes; reversed for second to fifth graders); The usefulness of induction and errorful learning varies with the type of terminal task being taught [+W] (conceptual analysis)
Learners are advanced on these items, or already accurate on the first unaided round Drop the mediators, cues and wide band; give mixed unaided items with correction only on misses Instructional guidance that helps novices can become redundant or counterproductive as expertise grows. [~M] (cognitive-load studies, carried to association and precision practice untested)
Learners are young children (child) Keep cues and mediators ready-made; correct every error overtly by showing and saying the right answer and having the child repeat it; short blocks (5–8 minutes) Overt correction of errors speeds young children's learning compared with knowledge of correctness alone [+W] (second-hand; block length is an untested proposal)
Learners are adults (adult) Correction can be brief (the right answer shown); invite them to build their own mediators; spend the time saved on unaided, mixed items Overt correction of errors speeds young children's learning compared with knowledge of correctness alone [~W] (adults did not learn faster with overt correction, reported second-hand)
Learners have mild learning or intellectual disabilities, or find remembering hard Teach the mediators explicitly (keyword plus picture, acrostic), present them for the learner rather than asking the learner to invent them, and fade the picture only after two sessions of correct recall Mnemonic Instruction Improves Recall For Students With Disabilities [+S] (recall of facts and word meanings; the fading rule is an untested proposal)
Learners have low confidence, or have given up after repeated failure Start with the band so wide that the first attempts succeed, narrow only after visible success, and keep errors private (feedback on the learner's own screen or sheet, not aloud) untested proposal; from the earlier page's concern about motivation loss when criteria tighten abruptly
Learners have low literacy, or are learning in a second language Use picture and sound mediators rather than written acrostics; check the learner can read or say the mediator before using it untested proposal (the mnemonic claim's own Discussion notes learners must be able to decode the mnemonic; not tested)
Goal is retention at a delay (a week or a term away), not performance at the end of the lesson Fade cues faster, accept some misses with immediate correction, and re-test unaided at the start of later sessions; do not judge readiness on cued items Initial retrieval conditions that provide less cue support, such as free recall rather than recognition or fewer letter cues, tend to produce better retention despite lower initial success [~M]; A review reports that retrieval practice effects grow more robust as initial retrieval success rises, especially above 75%, and separately that giving more retrieval cues yields smaller effects [~M] (both word-learning studies reported in a review chapter; keep most practice successful, not all of it trivially so)
Learners make errors with high confidence (they are sure of a wrong pairing) Ask for a confidence rating before showing the answer; correct confident errors at once and bring them back sooner. A confident error, corrected, is an occasion to fix the pairing, not only a cost to avoid High-confidence errors are more likely than low-confidence errors to be corrected on a later test after feedback. [~M] (a review and one experiment; qualifies "minimize errors" for recall items)
The brief reports an error learners actually made (a swapped pair, a habitual drift in timing) Use it as a contrast: show the wrong and the right response side by side once, then return to cued, correct practice on that item untested proposal; the comparison of wrong with right is modelled on Error Analysis
Practice items are multiple choice Show the correct answer after every item, including correct ones answered with low confidence Taking initial multiple-choice tests without feedback can lead students to later produce the incorrect lure answers they selected [+W]
Setting is online-self-paced, or phone-only Build the cue fading and correction into the items: each miss shows the right answer and returns the item later in the set; a short unaided check opens each session Feedback Enhances Retrieval Practice [+S] (classroom quizzing and a lab word-pair study, carried to self-paced apps untested)
Some learners join remotely while others are in the room Give everyone the same practice items on their own device or sheet, with the answer shown after each one; the teacher watches the shared error counts rather than individual voices untested proposal
One teacher with a large class Have all learners respond at once (cards, mini-whiteboards, a poll); correct the commonest error for the whole class, then re-present that item; work one-to-one only with those still missing it untested proposal; correction evidence is from quizzing and lab studies (Feedback Enhances Retrieval Practice [+S])
Facilitator time is scarce (volunteers, a busy shift, a tutor checking in for a few minutes) Put the cue fading and the answer check into the materials (answer on the back of a card, a self-check gauge, an app); use the facilitator's few minutes for narrowing the precision band and for confident errors untested proposal
Only one session, of fixed short length In 20 minutes: 2 minutes measuring current attempts, 5 minutes introducing 6–8 items with mediators, 10 minutes cued practice with fading and correction, 3 minutes unaided check. Report only the end-of-session result; claim no retention untested proposal
The brief names its own routine (its own drill, flashcard set, rubric or tolerance) Keep it; change only what the evidence speaks to: correct every miss at once, take cues away rather than adding them, narrow a tolerance gradually untested proposal for the routine; the correction step rests on Feedback Enhances Retrieval Practice [+S]
The brief gives its own categories, criteria or final tolerance Use those as the final criterion and the items; set only the starting band and the steps between them untested proposal
Material is private or sensitive (a learner's errors seen by peers, workplace records) Keep error records with the learner; feedback is shown only to the learner unless they choose to share it untested proposal
Adults in a work setting where precision matters for safety (a dose, a torque, a measurement) Practise on representative tasks with the real instrument; narrow to the job tolerance before unsupervised work; require the final tolerance on two separate occasions, unaided; never relax the final criterion, only the starting band untested proposal
A whole performance for a real audience (a recital, a handwritten letter, a demonstration) that the brief wants kept whole Shape precision inside short practice runs of the whole piece, not in isolated drills only; finish each session with one whole run judged against the current band untested proposal
Stakes are high, or the final check is unaided Judge readiness only on unaided, uncued responses at the start of a session; record every cue and correction given in practice untested proposal

Nine of the twenty-two rows rest on a claim, and two more (one teacher with a large class; the brief's own routine) cite one for a single step. Of the nine, one (mnemonics for students with disabilities) tests a part of this design with the learners named in the row; the rest carry a neighbouring finding (retrieval practice, productive failure, cognitive-load studies, second-hand reports) to this setting and say so. The other eleven rows are untested proposals.

Observation, state and explanation

What can be observed is a response under stated conditions: the item, the cue level shown, the current tolerance band, whether the response met it, how far off it was, the time taken, and what the learner did after a correction. "Has the association", "has the precision" and "is relying on the cue" are inferred states. A correct response with a cue shows what the learner does with that cue, not what they do without it.

Response Plausible explanations Observation that could change the interpretation
Correct with the cue, wrong as soon as it is removed The cue carried the answer; the cue was removed in too large a step; the mediator was never linked to the meaning Remove only part of the cue on the next presentation; ask the learner to say the mediator, then the answer.
The same pairing wrong across several sessions A competing pairing the learner already holds; the mediator is confusing or unfamiliar to this learner; corrections were not attended to Ask for a confidence rating before the answer; ask what the mediator reminds them of; try a different mediator for that item only.
Precision improves within a session, then is back to the old spread at the start of the next Little consolidation; the band was narrowed faster than the learner could keep; fatigue at the end of the previous session Start the next session one band wider and measure; compare the first five attempts of each session across a week.
Stops trying, rushes or refuses as the band narrows The criterion tightened too fast (the earlier page's warning); the task is no longer valued; public errors are costly Widen one step and see whether effort returns; ask the learner what makes the task worth doing; move feedback to private.
Accurate in practice, inaccurate on the real task (an instrument, a form, a performance) Practice items differ from the real ones; time pressure or an audience; the cue was present in practice Give one real-task attempt with no cue at the start of the next session and note the conditions.

These rows are proposals, untested with learners as diagnostics. A probe is itself practice and changes what it measures, so record each one and keep the explanations that remain open.

Evidence and qualifications

No claim in the wiki compares error-minimizing instruction (progressive tolerance, mediators and stimulus fading) with errorful training in human learners. The claims below test parts of it or its limits.

  • Mnemonic Instruction Improves Recall For Students With Disabilities [+S]: two syntheses read from abstracts. One pooled mnemonic-strategy studies with students with learning and behaviour problems and reports a mean standardized effect of 1.62, with laboratory and classroom studies alike; the other reviewed 20 studies with 669 secondary-age youth with disabilities and found improved recall of word meanings and facts. It supports the mediator half of this page for recall, in those learners; it does not test fading or error rates, and the claim page notes that the evidence is weakest for understanding and long retention.
  • Feedback Enhances Retrieval Practice [+S]: its table counts 8 distinct studies. In classroom quizzing, quizzes followed by corrective feedback showed a larger testing effect than quizzes without it (a between-study moderator; an earlier meta-analysis found no such moderation). In a word-pair experiment with 258 participants, the correct answer after a wrong response greatly raised retention a week later; feedback after correct responses made little difference. It supports correcting errors at once; it does not show that preventing them is better than correcting them.
  • Minimal guidance is less effective for novices than explicit instruction [+M]: its table counts 13 distinct studies across it and narrower claims. Across 580 comparisons explicit instruction beat unassisted discovery (d = 0.38), and discovery enhanced with feedback, worked examples or scaffolding did well; with 112 children, many more learned an experimental-design procedure by direct instruction. It is about guidance for novices in concepts and procedures, not about error rates, so it is carried here at M.
  • Errorless discrimination training avoids emotional responses and yields more effective transfer than errorful training (Terrace's pigeon studies) [+W]: Glaser's report of Terrace: pigeons trained on a discrimination by progressively reducing stimulus differences made few errors, and only birds trained with errors showed "emotional" responses, bursts of responding to the wrong stimulus, and weaker transfer. It is the only direct comparison behind this page, in animals, second-hand, with no effect size.
  • Overt correction of errors speeds young children's learning compared with knowledge of correctness alone [~W]: Glaser's review attributes to Suppes and Ginsberg (1962) faster learning with overt correction in young children's concept formation, and to Burke, Estes and Hellyer (1954) no gain from it in adults. It qualifies the correction step by age.
  • The usefulness of induction and errorful learning varies with the type of terminal task being taught [+W]: Glaser's conceptual analysis: precision and associations do not appear to need induction and errorful learning, induction may suit concepts, and for rules and strategies the paper is "less than definitive". It is the source of this page's scope, not a test of it. (The earlier page marked it [+M], above its cap.)
  • Self-discovery through errorful exploration is basically incompatible with error minimization in instruction [~W]: Glaser's argument that discovery sequences need blind alleys, so a sequence cannot both minimize errors and leave learners to discover. It marks where this page stops.

What limits it

Learners, tasks and outcomes differ across these claims (pigeons, students with disabilities, college students on word lists, school students on mathematics concepts), so do not rank them by effect label. None measures a precision outcome, and none compares progressive narrowing of a tolerance with a fixed final criterion.

The evidence by question

Further evidence, not yet read against this model

Objective and learner-valued goal

The designer's objective is usually an unaided response within a stated tolerance, or unaided recall of a set of pairings, at a stated horizon. Elicit separately what the learner wants: to play a piece at a concert, to read their own child's school letters, to pass a certification, to stop feeling foolish when they get it wrong. Agreement and divergence are both observations. A learner who values never being seen to be wrong may welcome error-free practice and avoid the unaided checks that show whether it worked; one who wants the real performance may find narrow drills pointless and stop trying.

For example, an adult learning a new musical instrument may want to play one song at a family event, while the teacher's objective is even timing on scales within a narrow tolerance. Shaping timing on the song's own phrases serves both; scales alone serve only the teacher's. Record both, so that a tight band on scales is not read as the capability the learner came for.

What would revise this model?

The expectation should weaken if comparisons of error-minimizing training with errorful training for people, on the same precision or association targets with aligned unaided outcomes, show no advantage, or an advantage at the end of practice that disappears a week later. If faster cue removal with more errors (and immediate correction) consistently beats near-errorless fading for delayed retention, as the retrieval findings suggest for word lists, the model should say "fade fast and correct", not "minimize errors". If a tolerance narrowed in one step works as well as one narrowed gradually, without loss of effort, the shaping half is a preference only. If high-confidence errors, corrected, turn out to be the best teaching moments for associations, minimizing them is the wrong goal for that case. Do not protect the model by calling every failure "too many errors allowed" after the fact.

End-of-practice accuracy, unaided performance at a delay, transfer to new items or the real instrument, and use in the setting the learner values are separate claims. The present evidence establishes no starting band, no step size, no error rate and no rule for when to remove a cue.

Examples

Key Sources

  • Glaser, Robert. (1966). Variables in "Discovery Learning." Learning Research and Development Center, University of Pittsburgh. https://eric.ed.gov/?id=ED010518
Page metadata
Field Value
type principle
id minimize-errors-for-precision-and-associations
title Teach response precision and new associations with minimal errors, using mediators and stimulus fading
description For a learner new to a precise response or a set of new pairings, keeping responses almost always correct (a tolerance narrowed in small steps, mediators linked to known items, cues faded across trials, every error corrected at once) is expected to give accurate unaided performance with fewer rehearsed errors than errorful trial and error, though no claim in the wiki tests this with human learners and cue support kept high can weaken retention.
status review
canonical True
generated by claude/unspecified on 2026-10-15
trust unverified — no one has confirmed this page's content
sources 1 — glaser-1966

Frontmatter is the page's machine-readable record — see the Schema & Guide.