Skip to content

Assessment for Learning

Principle · All principles
Evidence · 28 claims (13 for, 13 mixed, 2 against) · 34 studies (11 causal, 8 quant-synthesis, 5 review, 5 theoretical, 2 associational, 2 qualitative, 1 design), q1–q4 · 8 of 34 report an effect size · 21 claims rest on one study

Conditional relationship

Assessment for learning concerns who uses assessment evidence during instruction, against what criteria, and under what stakes. For learners working towards a capability they have not yet reached (in pre-k-12 or postsecondary settings; the evidence below does not report expertise), a response elicited before the outcome is judged is expected to support later performance on that capability only when three conditions hold: the learner and the teacher share the criteria the response is read against; the response is produced under conditions that make revealing partial understanding safe; and a revision, regrouping or reteaching step can follow before the summative judgement. The expected change is in general-achievement and, where learners judge their own work, in self-regulated strategy use and assessment-accuracy, at the horizon of the next aligned task or unit; the wiki holds no evidence for a longer horizon. One non-randomised study tests the third condition alone (instruction with and without a feedback-and-revision step, favouring the step at post-test and four weeks later); no claim tests the three conditions together or separates them.

The converted formative assessment principle and its pattern model the narrower teacher-side loop inside this relationship: a response is read against rival hypotheses and a contingent next action is chosen, then reobserved. This page owns what surrounds that loop: whether learners themselves can read the evidence (self-assessment against criteria, and how well calibrated it is), and whether the purpose and stakes of the checkpoint let the evidence be honest. Use the formative assessment pattern as the reusable policy; no separate assessment-for-learning pattern exists.

Fitting the design to a situation

Ask for three facts when the brief omits them: whether the evidence will be graded or counted, whether learners see the criteria before they attempt the task, and whether there is a revision step before the outcome is judged.

If the brief says… Then change… Basis
School classes (pre-k-12), with achievement as the goal Gather evidence during the unit and use it before the unit test; expect a small overall gain, larger where learners self-assess Assessment for learning improves achievement [+M]
Learners will judge their own work against criteria Show the criteria with an exemplar before the attempt, have learners self-assess a draft, then revise Self-assessment improves self-regulated learning [+M]; Rubric-referenced self-assessment improved work [~M]
Learners are novices, or the tasks are unfamiliar Train self-assessment with modelling examples before relying on it, and check self-ratings against scored work Self-assessment accuracy is low without training [~M]; Prior knowledge needed for accurate self-assessment [~M]
Assessment evidence will carry marks or be high-stakes Separate the formative evidence from the grade and lower the stakes on in-class checks; the one comparison in the wiki is confounded Low-stress team-based learning format scored higher [~W]; Formative assessment implementation barriers [~W]
Peers can review each other's work Add peer assessment; its effect is about the same as self-assessment and does not depend on a fixed rubric Peer and self-assessment improve performance equally [+W]; Peer assessment improves performance [+M]
The course has repeated drafts (writing, projects) Schedule a diagnose–revise step in each cycle and judge the outcome after revision Diagnostic review gave larger writing gains [+W]; Group-by-cycle interactions for formative outcomes [~M]
Pre-service teachers learning concepts such as the nature of science Embed formative checks with reflection in the explicit instruction, and recheck four weeks later Formative assessment with explicit-reflective instruction gave higher post-test scores [+W]; Higher retention four weeks later [+W]
Feedback will come from AI rather than the teacher AI feedback can match teacher feedback on a group project; prefer prompts that ask learners to reflect over handing them corrections AI and teacher feedback gave equal project gains [~M]; Reflective feedback beat direct GenAI feedback on delayed transfer [~M]; Rubric-based AI feedback scaffolded learning and reduced teacher workload, per the authors [+W] (authors' interpretation, untested)
English learners Use the same cycle, with criteria and feedback in language the learners can act on Teacher feedback moves learning forward [+W]
The assessment is a one-off entry or placement test Treat it as summative: tell learners beforehand what it covers and what it decides, and gather formative evidence elsewhere; a single sitting offers no revision step Students experience placement as a one-day isolated event with little advance information [-W] (describes experience, no outcome measured)
Competitive or norm-graded classroom culture Make self-ratings private and ungraded, and compare them with an ungraded task before drawing conclusions untested proposal

Observation, state and explanation

Record what was observed (a response to a named task, a self-rating, a revision, a silence) with its conditions: stakes, who sees the result, whether criteria were shown, response mode and help. Understanding, calibration and willingness to reveal uncertainty are inferred states; a correct answer, an accurate self-rating and a candid "I don't know" are each consistent with more than one of them. The rows below are proposals, untested with learners.

Response Plausible explanations Observation that could change the interpretation
Learner rates own work much higher than the teacher does Missing task knowledge to see the errors; criteria not understood; incentive to inflate under competitive or graded norms Ask the learner to point to evidence in the work against each criterion; repeat the rating privately and ungraded; compare with a peer's rating of the same work.
Few errors or questions appear on a checkpoint that carries marks Learners have reached the target; learners avoid revealing uncertainty because the mark counts; the task is too easy to discriminate Re-elicit the same capability on an ungraded, private item of comparable difficulty, and check whether the item separates learners at all.
Self-ratings become accurate but the work does not improve Calibration has improved without a revision opportunity; criteria are understood but strategy is missing; the rated task is not the target capability Check whether a revision step followed the rating, and whether a new unassisted task shows the same gain.
Results are collected but the next lesson is unchanged No usable next action was planned; the evidence did not discriminate between options; time or pacing constraints override it Record the decision that each checkpoint was meant to inform before it is given; compare what was planned with what was done.

A single contrast can shift confidence between explanations; it does not identify a cause. Do not read compliance, a grade or a confident self-rating as a learning state.

Evidence and qualifications

  • Assessment for learning improves achievement [+M]: syntheses of US K–12 (Lee et al.) and K–12 (Kingston and Nash) formative assessment studies with control conditions report small positive overall effects on achievement (d = .29; about .20). In Lee et al.'s meta-regression, interventions supporting student-initiated self-assessment show larger estimates (d = .61), as do formal formative evidence such as written feedback on quizzes (d = .40) and a medium cycle within or between units (d = .52). These feature estimates compare study configurations within a synthesis (synthesis-mixed grain, assignment per study unknown); the claim page says they are not randomised comparisons, not additive and not a forecast for a new course, and no common outcome horizon is reported. Partly checked against its sources.
  • Self Assessment Improves Self Regulated Learning [+M]: a meta-analysis of 19 experimental and quasi-experimental self-assessment interventions in K–12 and higher education reports a small effect on self-regulated strategy use (d = 0.23, 95% CI 0.08–0.39, 12 effect sizes) and a larger one on self-efficacy (d = 0.73); the other two SRL analyses have intervals crossing zero. It bears on the learner-side route of this model, not on achievement directly, and its page notes that gains in scores are thought to be mediated by strategy use. Not yet checked against its sources.
  • Self Assessment Accuracy Is Low Without Training [~M]: in a randomized 2×2 experiment with 80 Dutch secondary pre-university students solving genetics problems, those who studied modelling examples of self-assessment rated their own posttest performance more accurately than those who did not (ηp² = .10); a meta-analysis of 51 higher-education studies found self–teacher mark correspondence inconsistent, closer in advanced courses and better-designed studies. This limits the model: a learner's self-assessment is evidence a designer can act on only once its accuracy has been taught or checked. Not settled from the abstracts available.
  • Prior Knowledge Needed For Accurate Self Assessment [~M]: a synthesis of 22 meta-analyses reports moderate self–performance agreement (mean r = .29), higher for domain-specific judgements and familiar tasks; a semester-long classroom study of 96 undergraduates found the lowest performers grossly overconfident before and after exams. The associational evidence qualifies the learner-side route most for exactly the learners who most need revision. Not settled from the abstracts available.
  • Dental students in a revised low-stress TBL format scored higher on the individual summative assessment (iSAT) than medical students in the original high-stakes format, in a non-randomised comparison of two cohorts [~W]: the only claim found that contrasts stakes directly compares 119 dental students in a revised low-stress format with 161 medical students in the original, reporting a 5.2% higher iSAT (p < .05) and no effect size. Cohort, programme and format differ together (controlled-nonrandom), so it cannot separate stakes from population. Not yet checked against its sources.
  • Formative assessment with explicit-reflective instruction gave higher post-test scores [+W] and higher retention four weeks later [+W]: the wiki's one comparison of the same instruction with and without a feedback-and-revision step. In a pretest–posttest nonequivalent control group design with 33 pre-school prospective teachers per group, the group that designed and presented its own nature-of-science activities, received the instructor's feedback on them and revised them scored higher than an explicit-reflective-only group on the NOS scale (means 101.68 vs 92.47) and the SK scale (108.15 vs 100.71), and again on a retention test four weeks later (100.47 vs 90.59; 106.50 vs 99.15); no between-group effect size is printed. The groups did not differ at pre-test Pre-test equivalence [~W]; both gained from pre to post, with larger within-group effect sizes in the experimental group (r = .83 and .61, against .41 and .24) Both conditions gained, larger in the experimental group [+W]; and neither declined significantly from post-test to retention, which does not establish that scores held No significant decline to retention [~W]. It bears on the third condition (a revision step before the outcome is judged) but cannot separate the feedback from the designing and presenting that came with it, the groups were not randomised, and the outcomes are view scales on one study.
  • Contingent scaffolding improves learning more than fixed or absent support [+M]: bears on the step that follows the evidence. In one-to-one tutoring of long division, support adjusted to each response beat fixed, moderate or no support on immediate and one-month tests (Pratt and Savoy-Levine, N = 8 per condition), and a synthesis of 30 dynamic-assessment studies ranks contingent scaffolding above coaching but below explicit strategy training; the entries print no effect sizes. The claim page notes that most of this evidence is one-to-one and that contingency across a whole class is rare and hard. It supports making the next action depend on the evidence; it does not show that classroom checkpoints achieve that contingency. Not settled from the abstracts available.
  • Self-monitoring improves self-regulation and supports better learning decisions [+W]: a review (Zimmerman 2002) and a theoretical synthesis (Butler and Winne 1995) give the mechanism the learner-side route assumes: feedback becomes usable when learners compare their performance with a target and revise. No study on the claim page tests it, so it explains the route rather than supporting it empirically.
  • AI and teacher feedback gave equal project gains [~M]: 47 university student groups (N = 238) were randomly assigned to feedback from one of two AI models or an expert teacher; project performance rose from pre- to post-feedback in all three (rrb = 0.77), with no difference between sources and equivalence shown for one model. There is no no-feedback condition, so it compares sources of feedback, not feedback against none. Rubric-based AI feedback [+W] adds only the authors' interpretation that rubric-based AI feedback let teachers scaffold while reducing workload.
  • Students experience placement as a one-day isolated event [-W]: focus groups at five California community colleges describe assessment and placement as an isolated one-day event with minimal advance information, disconnected from recent work: the shared-criteria and revision conditions absent. It reports experience, not an outcome.

No claim in the wiki tests the stakes and goal-structure part of this model as stated: that grading pressure or competitive classroom norms make learners hide uncertainty or inflate self-assessments, or that a teacher's beliefs about ability change how assessment evidence is used. The previous body argued this from Dweck, Harlen and Patrick et al., which are listed under Key Sources but have no claim pages here. Treat that branch as a hypothesis.

The evidence by question

Further evidence, not yet read against this model

Claims read against the model on 2026-10-15 and found not to bear on its relationship, steps or situations; each says how far it bears and how far its own sources have been checked. The others once listed here are now in the sections above.

Objective and learner-valued goal

The designer's objective is a capability measured by a named instrument at a stated horizon; the learner's valued goal is what they want the work to achieve and why. Under assessment for learning the two often diverge on the checkpoint itself: a learner may value the mark, the designer the information. Ask, record both, and treat agreement as an observation.

For example, in a first-year writing course the designer's objective may be an argued essay scored on an analytic rubric at the final submission, while a learner's valued goal is "not to lose marks on the draft". If the draft is graded, that goal rewards hiding weak sections; if the draft is ungraded and the learner rates it against the rubric before feedback, the same goal can be served by revision. Which arrangement the learner experiences as safe, and whether their self-rating matches the teacher's, are things to elicit, not assume.

What would revise this model?

Revise it if comparable learners using criteria-based self-assessment with a revision step do not outperform comparable learners without one on an aligned delayed task under a defensible comparison; if removing marks from checkpoints does not change how much uncertainty learners reveal, measured by a private re-elicitation; or if trained self-assessment remains as miscalibrated as untrained for the learners in question, in which case the learner-side route should rely on external evidence. A synthesis-level average, a feature estimate from meta-regression, or a confounded cohort contrast does not settle any of these. Do not protect the model by relabelling every null result as poor fidelity.

Claims

Examples

Key Sources

  • Black, P., & Wiliam, D. (1998). Assessment and classroom learning. Assessment in Education, 5(1), 7-74. https://doi.org/10.1080/0969595980050102
  • Hawe, E., & Dixon, H. (2017). Assessment for learning: A catalyst for student self-regulation. Assessment & Evaluation in Higher Education, 42(8), 1181-1192. https://doi.org/10.1080/02602938.2016.1236360
  • Dweck, C. S. (2000). Self-theories: Their role in motivation, personality, and development. Psychology Press.
  • Harlen, W. (2006). On the relationship between assessment for formative and summative purposes. In J. Gardner (Ed.), Assessment and learning (pp. 103-117). Sage.
  • Patrick, H., Anderman, L. H., Ryan, A. M., Edelin, K. C., & Midgley, C. (2001). Teachers' communication of goal orientations in four fifth-grade classrooms. The Elementary School Journal, 102(1), 35-58. doi:10.1086/499692
  • Arduini-Van Hoose, N. (2020). Teacher's purpose and belief. In Educational psychology. Retrieved from https://edpsych.pressbooks.sunycreate.cloud. CC BY-NC-SA 4.0.
  • Mustafa Metin. (2022). Comparing Effects of Two Different Explicit–Reflective Instructions on Pre-School Prospective Teachers' View about Nature of Science and Scientific Knowledge. Journal of Science Learning, 5(1), 165-175. https://doi.org/10.17509/jsl.v5i1.33190
  • Kasten, Margaret; Meiring, Steve; Mikesell, Anne. (1997). Mathematics. Competency-Based Education Assessment Series. https://eric.ed.gov/?id=ED417935
Page metadata
Field Value
type principle
id assessment-for-learning
aliases formative-assessment-embedded-explicit-reflective-instruction, embed-assessment-in-ongoing-instruction
title Assessment for Learning
description Assessment evidence gathered during instruction is expected to support learning when teacher and learner read it against shared criteria, under stakes that let partial understanding show, with a revision step before the outcome is judged.
status review
canonical True
generated by claude/unspecified on 2026-10-15
trust unverified — no one has confirmed this page's content
sources 4 — black-1998, hawe-2017, mustafa-metin-2022, kasten-1997

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