Skip to content

Feedback Loops

Principle · All principles
Evidence · 31 claims (18 for, 11 mixed, 2 against) · 47 studies (13 causal, 11 review, 10 quant-synthesis, 5 design, 4 qualitative, 2 associational, 2 theoretical), q1–q4 · 8 of 47 report an effect size · 22 claims rest on one study

Conditional relationship

This page owns the loop as a whole: a learner's performance is observed, information about it reaches someone who can act on it (the learner, an instructor or a system), the next attempt is changed by that information, and the next attempt is observed again. For a learner working on a task where performance can be seen and repeated on comparable items (a procedure, a concept or principle applied to new cases, a draft of a complex-skill product such as writing), information about the performance is expected to improve later performance only when the loop closes: the information says what to change, a next attempt follows soon enough to use it, and that attempt is checked on a new item rather than the corrected one. Information that ends the episode (a score, a comment returned after the unit has moved on, praise of the person) has no route to an effect in this model. Over sessions the loop is expected to move from instructor- or system-run towards learner-run, as the learner learns to compare their own work with the target.

The relationship is conditional on what the information carries (corrective and high-information feedback are supported on average; feedback that draws attention to the self can make performance worse), on whether the next attempt actually uses it, on who can run the loop at the scale of the group, on the learner's expertise, and on the horizon measured (immediate-performance on the corrected item, near-transfer to a new item, delayed-retention). No claim in the wiki compares a closed loop with the same feedback given without a next attempt, or varies the number or length of cycles. The nearest evidence tests packages that contain a closed loop (correct-and-recheck mastery programmes, draft–feedback–revision peer review, contingent tutoring, intelligent tutoring), the content of feedback on its own, and self-monitoring as a mechanism. So the design below is a labelled default, not a tested model.

How the converted siblings sit beside this page, so a designer can go to the one that holds each decision: Formative Assessment holds how a single response is interpreted against rival explanations and which next action it warrants; this page holds whether that action happens, is re-observed, and repeats. Immediate Feedback holds when the information arrives relative to the attempt; this page only requires that it arrive before the next attempt it is meant to change. Assessment for Learning holds shared criteria and the stakes that make honest evidence possible. Mastery Learning holds one specific loop, a criterion-gated correct-and-recheck before advancing between units. Self-Regulated Learning holds the learner-run loop of planning and monitoring; this page hands over to it. Check-ins holds loops fed by learners' self-reports of their state. The formative assessment pattern and the mastery learning pattern carry reusable loop policies; their branches are untested proposals.

Default design, while the relationship is untested

The earlier page's guidance (observable performance, actionable information, an opportunity to respond, interpretation support; avoid delay, overload, low-information scores and permanent external correction), kept as a concrete default. Each step states its evidence.

  1. Plan the next attempt before the first one. For every task that will be observed, write the follow-up attempt in advance: a matched item for a step or problem, a revision of the same draft for a written or made product, a re-run of a performance. If no next attempt exists, the feedback is commentary. Untested proposal; from the earlier page ("opportunity to respond").
  2. Make the performance observable and aligned. Elicit a short response that shows the capability itself (a worked solution, a paragraph, a recorded attempt), not only a right/wrong answer, so the information can say what to change. How to read the response is modelled in Formative Assessment. Untested here as a step; in asynchronous online courses, the authors of one course audit argue that without an output that receives feedback there is little to show what a learner did (feedback makes behaviour seen [+W], an argument, not a comparison).
  3. Give information about the task and what to do next, one or two targets at a time. Say what was wrong or missing and the next move; leave out praise of the person. Evidence: feedback improves learning on average, more the more information it carries, and more than a third of feedback interventions in one synthesis made performance worse, the more so as they drew attention to the self (Feedback improves learning by a medium amount on average, but its effect varies widely with the information it carries, and a substantial share of feedback effects are negative [+S]); intelligence praise after success was followed by worse performance after a later failure in fifth graders (Praise for intelligence, rather than for effort, after success undermines children's motivation and performance after a later failure [-M]). The one-or-two-target limit is an untested proposal from the earlier page's warning about signal overload.
  4. Require a visible response to the information. The learner retries, revises and marks what they changed, or states the next step before attempting it. Evidence is thin: in one case study, required revise-and-resubmit after probing feedback moved one teacher candidate towards citing evidence, but progress was not linear and another candidate did not improve (probing feedback with revise-and-resubmit [+W]); in one study of 28 college students, those who complied more with a revise-summary prompt gained more (associational, revise-summary compliance [~M]); in log data from one platform, a workflow that had students select and prioritise AI-generated feedback and then discuss the selection was followed by more uptake (26.2%) than a directed workflow (14.1%) or a self-directed one (0.1%) (enacted feedback uptake [+M]; uptake is measured within each workflow, and no learning outcome is reported).
  5. Re-observe on a new item, not the corrected one. A corrected answer can be copied; the check that the loop worked is a fresh comparable item, unaided. For drafts, the check is the next piece of work, not the revised draft. Evidence for the recheck: correct-and-recheck mastery programmes outperformed conventional instruction on examinations in 108 controlled evaluations (Mastery Learning Improves Outcomes [+M]); the recheck itself is not isolated there.
  6. Change the support when the same error survives two cycles. Do not repeat the same message: switch to a worked example of the step, a contrasting case, or a demonstration, then a further new item. Evidence: support adjusted to each response beat fixed or no support in one-to-one tutoring of children on long division, at immediate and one-month tests (Contingent scaffolding [+M]); the two-cycle threshold is an untested proposal. The mastery claim's own discussion says recycling learners through the same material mostly re-exposes them to failure.
  7. Ask for confidence, and loop back to confident errors. Before showing the answer, ask how sure the learner is; correct confidently wrong answers explicitly and put the item back into a later session. Evidence: corrected high-confidence errors are better retained than low-confidence ones (hypercorrection [~S]); the claim's evidence is mostly recall of facts.
  8. Hand the loop over across sessions. A default arc for a unit of several weeks (untested proposal):
  9. Early sessions: instructor- or system-run loops on short items. Three to five cycles (attempt, information, matched retry) per new step in a session, each cycle a few minutes. In one observation study, secondary mathematics teachers trained in group-based mastery learning gave about 20 percent of class time to the corrective loop, having given virtually none before (trained teachers' loop time [+M]); use that as a planning figure, not a tested dose.
  10. Middle sessions: the learner checks their own attempt against criteria or a worked answer before receiving feedback, and the feedback then checks the learner's judgement as well as the work. Evidence for self-monitoring as the mechanism is theoretical and review-level (Self-monitoring [+M]); learners' self-judgements need checking (see Assessment for Learning).
  11. Late sessions: the learner runs the loop; external information comes only on unaided samples and at a delayed recheck about a week after the step was learned.
  12. Move on, or not. Move a learner on from a step when two consecutive new items are correct unaided and the delayed recheck holds; send a learner back into instructor-run loops when a delayed recheck fails. Untested thresholds; set locally and record them. How a criterion gates advancement between units is modelled in Mastery Learning.
  13. Close the instructor's loop too. After each checkpoint, record the decision it informed (reteach, regroup, extend, move on) and check at the next checkpoint whether that decision helped. Untested proposal; from the earlier page (an exit ticket that changes the next day's grouping). The step is not automatic: a brief citing earlier research reports that teachers who can identify where a learner is do not necessarily know the next instructional step (noticing is not enough [~W], reported second-hand), so plan the candidate next moves for each likely response before the checkpoint.

Fitting the design to a situation

The three facts that most change the decision:

  • Who can read each performance and how soon the next attempt follows. One-to-one tutoring, a teacher with thirty learners, peers, software and the learner alone close the loop at very different rates; most evidence for responsive loops is one-to-one. If unstated, ask: after an attempt, who will look at it, and when will the learner make the next attempt?
  • What the next attempt is. A matched item minutes later, a revised draft days later, and a performance next term make different loops. If unstated, ask: what will the learner do differently next time, on what task, and how will anyone see whether they did?
  • Whether the learner can judge their own work yet. A novice needs the information supplied; a learner who can compare their work with the target can run the loop with occasional checks. If unstated, ask: shown their own attempt and a correct one, can the learner say what differs?
If the brief says… Then change… Basis
Learners are novice on the task Run the loop for them: task-level information on every attempt, a matched retry at once, the step re-modelled after two failed cycles Contingent scaffolding [+M] (children, one-to-one, long division)
Learners are advanced, or already accurate unaided Shorten external loops: ask for self-evaluation first, give information only on what the self-evaluation missed, and loop on harder or transfer items Expertise reversal [~M] (review of guidance studies, carried to feedback; no crossover point given)
Learners are child (primary age) Shorter cycles with the adult reading each attempt; praise the strategy or the work, never ability; small groups so attempts can be seen Praise: Praise for intelligence, rather than for effort, after success undermines children's motivation and performance after a later failure [-M] (fifth graders). Cycle length: untested
Goal is a written or made product (complex-skill) Loop over drafts: draft, structured peer or teacher feedback against criteria, revision within a few days, then a new piece checked for the same targets Peer Feedback Improves Writing [+M] (higher education, 24 studies; against no feedback); revision interval untested
Goal is spoken second-language accuracy Correct errors in the exchange itself, with prompts that ask the learner to reformulate rather than only recasts; keep it up over many sessions Corrective feedback on L2 errors [+S] (two meta-analyses; gains held on delayed tests; prompts larger than recasts in the classroom synthesis; the two disagree on whether longer treatments do better, and younger learners benefited more, so the pooled effect fits adults least)
classroom, one teacher and a large group Collect every learner's attempt at once (mini-whiteboards, a poll, a short written item), give whole-class information on the commonest error, have everyone do one matched item, then go to those who still fail; use peers with a rubric for drafts; budget class time for the loop Trained teachers' loop time [+M] (secondary mathematics, time allocation only, no learning outcome); rubrics and peer feedback [~M] (rubrics changed peer comments, one study found less reflective feedback)
Little facilitator time or equipment: volunteers, busy shifts, no devices Run the loop in buddy pairs: one partner watches an attempt against a posted checklist of two or three points, names one thing to change, and the other retries at once; a wall tally records which checklist points were met; the facilitator rechecks a sample (one new unaided attempt from two or three people per shift) and changes the checklist if the same point keeps failing. No paperwork beyond the tally Untested proposal; peer reading carried from peer feedback on writing [+M], which tested drafts, not practical tasks
online-self-paced, no teacher present Build automatically checked items with a retry on a new item after each piece of information; route a learner who fails twice to a worked example; keep a delayed recheck in a later session Intelligent tutoring systems improve achievement over non-individualised instruction [+S] (intelligent tutoring beat large-group, textbook and other computer instruction, not human tutoring or small groups; gains were smaller on standardized than on locally developed tests)
Feedback comes from a generative AI tool Keep a person in the loop to help learners interpret and apply the AI's comments; have the learner choose and rank which comments to act on before revising, and check the revision against the chosen targets Feedback quality predicts revision only with teachers [~M] (one essay study: teacher feedback quality predicted revised-draft quality, AI feedback quality did not); enacted feedback uptake [+M] (uptake, not learning)
Mixed modes: some learners remote, some in the room, or phone-only Use one submission channel everyone can reach (a shared form or chat poll), return information in that channel, and give remote learners the same matched retry; for phone-only learners keep items short and one target per message untested proposal
One session only Run at least two complete cycles inside the session and end with one unaided new item; report what that item shows and claim no retention untested proposal
Stakes are high, or the work is graded Keep loop items ungraded and separate from the graded task; judge readiness only on unaided new items; let a revision replace an earlier mark if revision is the point untested; Assessment for Learning holds the stakes model (retakes and revisions is a recipe, untested)
Learners have low confidence or low literacy Give information privately, about the task, in short spoken or shown form (a demonstration, a marked example) rather than long written comments; one target per cycle untested proposal; person-focused praise is the one thing with evidence against it (Praise for intelligence, rather than for effort, after success undermines children's motivation and performance after a later failure [-M], children)
Adults in work settings (adult-workplace, workplace-clinical) Loop on real work outputs (a case note, a call, a procedure); a supervisor or peer reads one output per cycle against a stated criterion, and the next output is the re-observation untested; in one teacher-education case, probing feedback with required resubmission helped one candidate and not another (probing feedback [+W])
A product or performance for a real audience Run the loops in rehearsal with a critic using the criteria, before the audience; treat the audience's response as information for the next product, not a judgement on this one; do not show the audience the drafts or the critique if it should see only the finished work untested proposal
Learners receive feedback but do not use it Make the use visible: the learner marks what they changed, or answers "what will you do next?" before the next attempt; no new information until the last was used revise-summary compliance [~M] (associational, 28 college students, one tutoring system)
Learners answer confidently and wrongly (misconceptions) Ask for confidence before revealing answers; correct confident errors explicitly and bring the item back later hypercorrection [~S] (mostly recall; carried to misconceptions untested)
Information can only come later (marking takes days) Keep the loop closed by attaching a task: the learner uses the returned comments on a new item before the next unit, rather than reading and filing them; whether earlier would be better is not settled For one outcome, immediate and delayed feedback were equally effective (feedback after multiple-choice tests [+M], reported second-hand); timing is held on Immediate Feedback

Thirteen of the seventeen rows cite a claim, two of them (low confidence or literacy, work settings) only for a limit or a neighbouring case; four are untested proposals. Most cited claims are carried to the row's setting from another one (the contingent-support evidence is one-to-one with children, the hypercorrection evidence is recall of facts, the loop-time study measured teachers' time, not learning), and each row says so.

Observation, state and explanation

What can be observed in a loop is a sequence: an attempt under stated conditions, the information given (its content, source and delay), what the learner did with it, and the next attempt on a stated item. "Understood the feedback", "has corrected the error" and "can now monitor their own work" are inferred states. A correct retry on the corrected item shows the information was followed; it does not show the change carried to a new item or will last.

Response Plausible explanations Observation that could change the interpretation
Information given, next attempt unchanged The information was not read or not understood; it named a problem but not a next move; the learner lacked the skill to act on it; the next attempt came too late to connect Ask the learner to say what the feedback asks them to do, then give a matched item with that move prompted; compare with an item where the move is shown.
Correct on the retry of the corrected item, wrong on a new item The answer was copied; the correction was item-specific; the new item differs more than intended Give a second new item that varies only the surface, and ask for the step and why.
Same error survives two or three cycles A misconception the information does not address; the information is accurate but too abstract; the learner is anxious and attending to being wrong, not to the task Present a contrasting case where the misconception and the correct rule predict different answers; give the information privately and as a demonstration.
Learner revises only what was pointed out, and new errors of the same kind appear Compliance with comments rather than a changed criterion; the criterion is not understood Ask the learner to mark one further instance of the issue in their own work before the next feedback.
Accurate at the end of a run of loops, weaker a week later Loops supported performance without consolidation; items too close together; interference Recheck at a delay on a new item and record every practice and feedback episode in between.
Learner's self-check agrees with the teacher's on some criteria and not others Calibration differs by criterion; the criteria are understood unevenly Keep external information on the criteria where the self-check misses, withdraw it where it agrees.

These rows are proposals, untested as diagnostics. A probe is itself another cycle and changes the state it measures, so record it, and keep rival explanations open.

Evidence and qualifications

No claim in the wiki compares feedback followed by a required next attempt with the same feedback given without one, or varies the number, length or ownership of cycles. The claims below test neighbouring parts of the loop.

Learners, tasks, settings and outcomes differ across these claims, so do not rank them by effect size. None measures delayed-retention of a closed loop against an open one.

The evidence by question

Further evidence, not yet read against this model

Claims this page has read and does not use in the model above, with how far each bears on it.

Objective and learner-valued goal

The designer's objective is usually unaided performance on new items of the target kind, at a stated horizon, and, later, the learner's ability to run the loop themselves. Elicit separately what the learner wants from the feedback: a mark, reassurance, to finish, or to get better at something they care about. These diverge often in loops. A learner who values the mark may revise only what was pointed out; one who values not looking wrong may avoid the retry, or the confidence question. Agreement and divergence are both observations, and they change how the information should be framed and who should see it.

For example, a nurse in clinical onboarding may want to stop being corrected in front of patients, while the designer's objective is accurate medication calculations under time pressure. A loop that corrects calculations on paper in private, with a matched retry and a supervisor's check on the next real calculation, serves both; the same corrections given on the ward serve the objective and work against the learner's goal. Record both, so that a learner's avoidance of retries is not read as a lack of competence.

What would revise this model?

The expectation should weaken if comparisons of the same feedback with and without a required next attempt, in comparable learners on aligned new items, show no advantage, or show that the advantage at the end of the session disappears at a delay. If learners given high-information feedback improve as much without a re-observation step as with one, the recheck is a measurement convenience, not part of the mechanism, and the page should say so. If learner-run loops (self-checking against criteria) never match instructor-run loops for learners at a given stage, the hand-over should be later or dropped for them. If one teacher cannot run closed loops with a whole class however the time is budgeted, the model is a one-to-one or software design and should be stated that way. Do not protect the model by calling every null result "the loop was not really closed" after the fact.

Following the information on the corrected item, unaided performance on a new item, delayed retention, and the learner's own monitoring are separate outcomes. The present evidence establishes no number of cycles, no cycle length and no hand-over criterion.

Claims

Examples

Key Sources

  • Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81-112. https://doi.org/10.3102/003465430298487
  • Clark, I. (2012). Formative assessment: Assessment is for self-regulated learning. Educational Psychology Review, 24(2), 205-249. https://doi.org/10.1007/s10648-011-9191-6
  • Pollard, V., Armatas, C., & Pepe, A. (2025). Feedback is integral: Using a revised ICAP Framework to achieve active learning in an asynchronous online course. Online Learning, 29(3), 236-253. https://doi.org/10.24059/olj.v29i3.4555
  • Baker, A., Weisgrau, J., & Bristal Philyaw, K. (2022, May). Feedback loops: Mapping transformative interactions in education innovation. Digital Promise. https://doi.org/10.51388/20.500.12265/155
  • Merueña, J. A. A. (2023). Dynamic Language Learning Framework (DL2F) Instructional Model. EPRA International Journal of Research and Development (IJRD). https://eric.ed.gov/?id=ED637673
Page metadata
Field Value
type principle
id feedback-loops
aliases design-feedback-on-external-outputs-asynchronous, prefer-feedback-loops-over-feedback-systems, dl2f-continuous-improvement-principle
title Feedback Loops
description For a learner whose performance can be observed, information about it is expected to improve later performance only when the loop closes (the information says what to change, a next attempt uses it, and that attempt is checked again on a new item); corrective feedback and correct-and-recheck programmes are supported on average, but no claim here tests a closed loop against the same feedback without a next attempt.
status review
canonical True
generated by claude/unspecified on 2026-10-15
trust unverified — no one has confirmed this page's content
sources 5 — hattie-2007, clark-2012, pollard-2025, baker-a-weisgrau-j, merueña-2023

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