Skip to content

Personalization

Principle · All principles
Evidence · 80 claims (40 for, 34 mixed, 6 against) · 83 studies (28 review, 21 causal, 9 theoretical, 8 qualitative, 7 associational, 5 quant-synthesis, 5 design), q1–q4 · 7 of 83 report an effect size · 75 claims rest on one study

Conditional relationship

This page owns the question of what to tailor to each learner, and on what basis, when a group of learners shares one goal. Where learners differ in what they already know on the target task, in what interests them, in how they can access the material (a disability, a communication system, a language) or in what they prefer among equivalent means, tailoring the pace, the support, the context of problems and the means of working to that learner, while the goal, the criterion and the final assessment stay common, is expected to raise engagement and interest (motivation-affect, persistence) and, where the tailoring rests on a diagnosis of the learner's responses, achievement on an aligned measure (general-achievement, learning-time, mostly immediate). Tailoring on a learner's self-reported "learning style" is expected to add nothing.

The relationship is conditional on four things: what the tailoring rests on (responses on the task and stated access needs are better grounds than a style inventory or a guessed interest); what is tailored (support and pace for knowledge differences; context and choice for interest and motivation; means for access and preference); whether the learner's own account is asked for, since a designer's guess at a learner's interest often misses; and whether the tailoring is actually carried out across the weeks, not only adopted.

No claim in the wiki tests personalization as a whole against uniform instruction. One randomized experiment compares a diagnosis-based personalized path with a fixed sequence over the same content (undergraduates, probability theory). A meta-analysis compares intelligent tutoring systems with non-individualized instruction, and another compares AI-enabled personalized STEM teaching with teaching without AI (32 trials, g = 0.455), which tests the technology and its tailoring together. The one reported comparison of an adaptive system with static curricula is an institutional case study described second-hand in a review, with no assignment or effect size, so it is not counted as a test. No claim here tests interest-based problem contexts against standard contexts on achievement; the one randomized field experiment on a related move (localizing character names) found gains only in some schools. The choice evidence is about autonomy-supportive versus controlling conditions and rewards, not about choice of task as such.

How the converted neighbours sit beside this: Adaptive Learning owns the narrower, system-driven case, choosing each next task, hint or check from a learner's recent responses, and its default design and evidence on tutoring systems, learner models and fidelity apply here unchanged; this page does not repeat them. Autonomy owns giving learners choice and voice; this page uses choice only as one thing that can be personalized. Mastery Learning owns the gate-correct-recheck rule that lets learners progress at their own rate. Flexible Grouping owns regrouping learners by current need, the classroom alternative to one-to-one tailoring.

Default design, while the relationship is untested

The earlier page's guidance and the recommendations of the two pages folded into it, kept as a concrete default a designer can act on and revise. Each step is a design proposal; its evidence status is stated beside it.

  1. Fix what stays common before deciding what varies. Write the unit's goal, the criterion for success and the final assessment once, for everyone; personalize how and when (path, pace, support, context, means), not what counts as mastery. Untested proposal; from the earlier page's requirement that shared goals keep outcomes comparable.
  2. Find out where each learner is, at the start of each unit. Give a 10–15 minute diagnostic on the unit's prerequisite skills (a short set of items scored by part, not a total), and ask each learner, in their own words, what they are interested in or what they use outside school. Dose is an untested proposal. Diagnosis-based paths beat a fixed sequence in one experiment: In a randomized experiment, personalized paths reduced learning time by 22.0% and improved post-test scores versus a fixed sequence [+M] (undergraduates, one session's content). Asking rather than assuming: Students frequently reported that personalized problem topics did not match their interests [-W]. Re-run the diagnostic after a break in schooling rather than carrying old placements forward, as reopening guidance for students with disabilities asserts without data: New baseline data are needed to support instructional planning for individual students and student groups after disrupted schooling [+W]. Read a low diagnostic score from older pupils with care, and look for rapid guessing before placing them on a slower path: on a remote, low-stakes adaptive test, Student disengagement on an adaptive interim assessment gradually increased across grade level [~W] (an adaptive test in spring 2020, not adaptive instruction).
  3. Tailor support and pace to the diagnosis, and withdraw support as performance grows. Learners who cannot begin get worked or partly worked examples and closer items; learners who succeed unaided skip ahead to harder or transfer items. Check on unsupported items before removing support. Instructional guidance that helps novices can become redundant or counterproductive as expertise grows. [~M]. For a novice, start the sequence with a worked example rather than a problem: Example-problem sequences reduce cognitive load and improve learning outcomes. [+M] (secondary-school novices, circuit troubleshooting, immediate test; problem-first pairs did no better than problems only). Adjust support to each response rather than fixing it in advance: Contingent scaffolding improves learning more than fixed or absent support. [+M] (mostly one-to-one tutoring; in the dynamic-assessment meta-analysis it beats coaching but not explicit strategy training; Adaptive Learning owns it). Which way difficulty moves matters as well as whether it moves: In one quasi-experimental VR shooting study (n = 50), reported in a review, rule-based dynamic difficulty outperformed progressive difficulty [~W] (one VR shooting study, n = 50, reported in a review; dynamic beat progressive difficulty). The response-by-response version of this step is modelled in Adaptive Learning. Where a teacher, not software, chooses the easier or harder material, base the choice on the learner's diagnostic and check results and write the basis down, because judgement can track a group label rather than the work: Preservice teachers chose the easy video significantly more often for an ethnic minority student than for an ethnic majority student, all other information being equal [-M] (a vignette experiment with 108 preservice teachers; 82% against 11% chose the easy video for an otherwise identical description; a choice of material, not a learning outcome). In the same article's second experiment, which also varied the student's achievement, achievement decided the choice and ethnicity did not: Student achievement level, not ethnic background, predicted preservice teachers' text-versus-video material choice in Study 2 [~M] (177 preservice teachers). A recorded result to tailor on is what that second experiment had and the first did not.
  4. Let each learner advance at their own rate through a sequenced continuum. Lay the unit out as an ordered series of short sections (about one lesson each), each ending with a short unaided check; a learner moves to the next section on passing it, and others get correction and a recheck. This works in whole-class, small-group or individual settings and does not need a one-to-one ratio. Untested proposal; the recommendation of the folded continuous-progress page (a 1974 curriculum guide, no claim cited). Where a learner is behind, the continuum should reach back below grade level along a learning progression: Teachers who made productive adaptations were more likely to consider learning progressions and adapt lessons to below-grade-level content [+W] (an association among teachers' adaptations, not a test of learner outcomes). Learners who start furthest behind are not the ones to hold back from a personalized path: In mathematics, personalized-learning students in the lowest baseline quintile were the most likely to out-grow their matched comparison students [+W] (schools matched to virtual comparison groups; a share of students out-growing their comparison, no effect size; mathematics only). The gate is modelled in Mastery Learning.
  5. Offer bounded choice among options that serve the same objective. Two to four tasks or media of equal rigour (a choice board), with the criterion shown on every option. Give novices and younger learners fewer options. Choice and autonomy support are expected to raise motivation, not achievement directly: Self-determination theory holds that autonomy support raises intrinsic motivation, and a meta-analysis finds expected tangible rewards undermine it [+M]; moderate rather than full agency, in one game-based study reported second-hand: Moderate learner agency in game-based learning yields better outcomes than no agency or high agency [~W]. A choice alone may do little; support that answers the learner's state does more: Robot adaptiveness effects on human autonomy are mixed: autonomy-supportive behaviours and as-needed assistance help, but mere choice or high interactivity may not [~W] (a human–robot interaction review; perceived autonomy, not learning). The number of options is an untested proposal.
  6. Set problems in contexts the learner named, and check that they fit. Rewrite (or generate) practice problems in the contexts learners gave in step 2, keeping the mathematics or structure unchanged; ask learners after a few problems whether the context is one they care about, and let them change it. Expert guidance recommends it: An IES practice-guide presentation recommends personalized problem contexts, asserting without cited studies that students engage more actively and perform better in them [+W]. Limits: assigned topics often missed students' interests (topic mismatch [-W]), and localizing names helped only in some schools (Localizing word-problem character names improved MATHia performance only for specific subgroup-majority schools, with no improvement in sense of belonging overall [~M]). Some topics do not fit the mathematics, and a generated rewrite then trades the structure against realism: The three refinement strategies make distinct trade-offs when personalizing problems to topics that resist natural alignment with the original math [~W] (one LLM-personalized problem, a case analysis). Check every rewritten problem against the original's mathematics before it is used, and offer a neutral context where the learner's topic does not fit.
  7. Where learners use different means to the same end, assess preference with the individual. For a communication system (speech-generating device, picture exchange, manual signs), an assistive tool or a format for access, offer the options in turn over several sessions, record which the learner chooses when free to choose and how quickly they acquire the target skill with each, and decide per learner, not by system type. Individuals with developmental disabilities tend to prefer speech-generating devices over picture exchange and manual signs, with substantial individual differences [+M]. The folded AAC page also reported that requesting skills were acquired faster with a preferred system for some children; the claim page records preferences, not that acquisition result. The number of sessions is an untested proposal.
  8. Do not tailor on learning styles. Do not sort learners by a visual, auditory or kinesthetic inventory and match the format to it; choose the format the content needs, for everyone. Learning Styles Matching Does Not Improve Learning [-S]; one of the few tests with the needed design, reported in the same review, found no match effect among medical residents: Cook et al. (2009) found no support for sensing/intuitive style-by-sequence interaction in medical education [-M] (123 internal medicine residents, problem-first against content-first modules; no effect size reported). In one media experiment series, format rather than a preference match decided who did better: In one author's classroom media experiments, intuitive learners tended to score above sensing learners, though not in every format comparison [~W] (sensing learners with voice and diagrams outscored intuitive learners with text only; a conference report, sample sizes not given). For the same reason, do not match a model, character or example to a group category assumed for the learner (gender, ethnicity, background) and expect learning to rise: in one experiment with a video model, Model gender and observer gender do not affect learning outcomes or near transfer from video modeling examples [-W] (a 2 × 2 experiment crossing model and observer gender, about 160 analysed; no main effect and no interaction on learning or near transfer, equivalence not tested), and matching character names to schools' majority groups helped only in some schools (localized names [~M]). Where a learner wants a context or model that reflects them, let them name or choose it (step 6); that is the learner's own account, not a group label.
  9. Review the tailoring weekly, and check it was enacted. Ten minutes a week: who advanced, who has been held at one section for two or more weeks, who chose the same easy option every time, and whether anyone has been routed around content the final assessment needs. Also check whether the easier options or slower paths have gathered in one group of learners whose check results do not differ from their peers'; if they have, the assignment followed an expectation rather than the work, and those learners should be re-placed from their results (easy video for the minority student [-M], above). Judge the personalization only where it was used. Personalized Learning Effects Vary With Fidelity [~M]; the weekly dose is an untested proposal. Judge it on the common check, not on a change in what learners do: in one study, prompts tailored to each learner's detected study behaviour changed that behaviour and not the test score (Personalized scaffolds increased frequencies of high cognition and monitoring relative to control, but learning outcomes did not differ between conditions. [~W]).

Dose and progression across sessions (untested proposal). Early sessions in a unit: the teacher sets each learner's path from the diagnostic, choices are between two options, and problem contexts are drawn by the teacher from the learner's list. Middle sessions: learners choose among three or four options and move through sections at their own pace with a self-check at the end of each. Late sessions: learners propose their own context or topic and plan the remaining sections, with the teacher reviewing the plan. A learner moves on from a section after passing its unaided check (for example, four of five items), and moves from teacher-set to self-chosen paths once they have passed two consecutive sections on the first attempt. Re-run the diagnostic at the start of each new unit, since expertise is specific to the topic.

When a condition is not met: if there is no time for a diagnostic, personalize only context and choice (steps 5–6) and keep the path common; if the options cannot be made equal in rigour, do not offer a choice; if no one will review the tailoring (step 9), keep the number of separate paths to two or three that a teacher can track.

Fitting the design to a situation

The three facts that most change the decision:

  • What the learners differ in, on this task. Differences in prior knowledge call for tailored support and pace; differences in interest for tailored contexts and choice; differences in access for tailored means. If unstated, ask: on a short diagnostic, how far apart are the learners, and do any need a different way of taking part?
  • Who or what will do the tailoring, and how often they can look. Software can tailor every item; a teacher with thirty learners can tailor sections and groups weekly; a volunteer who checks in briefly can tailor one thing. If unstated, ask: who sees each learner's work, and how often?
  • Which outcome counts, and whether it is common. Engagement, achievement on an aligned test, time saved and a learner's own goal move separately, and a personalized path is only comparable if the final criterion is shared. If unstated, ask: what must every learner be able to do at the end, and how will it be checked?
If the brief says… Then change… Basis
Learners are novice on this task, or advanced / already succeed on the first item Novice: more worked steps and closer items, fewer choices. Advanced: skip guided sections after an unaided check and go to harder or transfer items Instructional guidance that helps novices can become redundant or counterproductive as expertise grows. [~M] (review of cognitive-load studies; gives no crossover point)
Learners are young children (child), or adult Child: two options at most, chosen by the teacher from what the child said; personalize contexts more than paths; in kindergarten, expect a small gain from AI-driven personalized practice and do not let it replace the teacher's tailoring. Adult: ask what they already do and want; expect a wider spread of experience, and offer more paths and more choice Adult: Adult learner groups are more heterogeneous than child groups, calling for individualized instruction and varied learning strategies [+W] (essay, no data). Child: School level significantly moderates effects: AI-enabled personalized STEM education is more effective in junior and senior high school than in kindergarten [~M] (a subgroup comparison across RCTs in one meta-analysis: kindergarten g = 0.118 against 0.42–0.53 from elementary to senior high; between studies, not a test within one); the number of options is untested
Learners have developmental disabilities and use or are choosing an AAC system Run a preference assessment with each child across the candidate systems over several sessions; choose per child, and re-check when the environment's demands change Individuals with developmental disabilities tend to prefer speech-generating devices over picture exchange and manual signs, with substantial individual differences [+M] (overview of reviews; whether preference tracks a child's characteristics or environment is unclear)
Learners have disabilities with different access needs (an AI or human tutor responds) Tailor two things separately: how much support (from the learner's baseline accuracy) and what style of support (from the access need, e.g. shorter steps, visuals, read-aloud) Prompt style is the largest driver of disability-aware Fit gains, and the Thinking Reward becomes effective only in combination with adaptive prompting [+W] (simulated learner personas, LLM tutor ratings, not learners' outcomes); Outcomes of students with disabilities improve when they receive specially designed instruction responsive to assessment data [+W] (brief citing four studies)
The brief proposes matching format to learning styles or a preference inventory Drop the matching; choose format by what the content needs; if learners differ, diagnose prior knowledge instead Learning Styles Matching Does Not Improve Learning [-S]
The brief proposes matching models, characters, names or examples to learners' group (gender, ethnicity, culture, background) Do not expect the match to raise learning, and do not assign a learner's contexts or "style" from a group label; keep one well-designed model or example for everyone, and where learners want to see themselves in the material, let each learner name or choose the context or model Model gender and observer gender do not affect learning outcomes or near transfer from video modeling examples [-W] (one 2 × 2 experiment; null, equivalence not tested); Localizing word-problem character names improved MATHia performance only for specific subgroup-majority schools, with no improvement in sense of belonging overall [~M]; assigned topics often missed: topic mismatch [-W]
Goal is engagement or interest in a subject (e.g. mathematics word problems) Use learner-named contexts with the structure unchanged; let learners change the context; measure interest directly rather than assuming the context raised it An IES practice-guide presentation recommends personalized problem contexts, asserting without cited studies that students engage more actively and perform better in them [+W] (practice-guide assertion); The generative AI relevance intervention did not significantly increase students' interest in the math class [~W] (high school, a relevance intervention, null on interest)
Setting is online-self-paced, with a platform or courseware Use a platform that diagnoses parts of the topic and orders each learner's path; expect time saved and a modest aligned gain; pair it with an instructor learners can reach In a randomized experiment, personalized paths reduced learning time by 22.0% and improved post-test scores versus a fixed sequence [+M]; ITS significantly outperform large-group human instruction, individual computer-based instruction, and textbooks or workbooks, but not small-group or individual human tutoring [+M]; Students in focus groups were positive about adaptive courseware but reported alignment, access, and reliability concerns [~W]; Courseware can bridge gateway math learning gaps through tailored tutorials, study tools, diagnostics, and targeted practice [+W] (faculty and student interviews; perceptions, not outcomes)
An AI tutor, chatbot or detection system does the tailoring Give it the learner's mastery record for the skill and its prerequisites, and keep a teacher able to override its task choices; have it plan the path step by step from the learner's responses rather than generate the whole path at once; if it reads the learner's emotional state, give it the teaching move to make for each state, not the state alone. If it tailors prompts from detected learning behaviour (what the learner has already planned, monitored or read), log which options each learner was shown to confirm the content really varied, and judge it on the learning measure, since behaviour can change while scores do not. Let each learner set the role it plays (study partner, coach, tutor) and how insistently it prompts them, rather than fixing one manner for all Role and insistence: Participants required GSR self-knowledge of an assertive, conscientious, friendly personality with a customizable role [~W] (interviews with university participants about a social robot; what they wanted varied from person to person; preferences, not outcomes). Real-time detection of SRL processes can personalize scaffold content, with all five scaffolds personalized to some extent but to varying degrees (scaffolds 2 and 5 most, scaffold 3 least). [+W] (counts for 18 learners in the personalized condition; shows the content varied, not that it helped); Personalized scaffolds increased frequencies of high cognition and monitoring relative to control, but learning outcomes did not differ between conditions. [~W] (same study, a summary of earlier analyses; null on learning, equivalence not tested); Removing explicit mastery information from learner profiles mainly degrades Adaptivity [+W] (model-generated paths, not learners); One-shot path generation raises Validity but lowers Adaptivity and Efficiency relative to step-by-step interaction [~W] (model paths scored by metrics, not learners); Emotion features improve tutor performance only when grounded in education-theory-based pedagogical strategies; emotion information alone misguides a smaller model [~W] (a prompt ablation on one model, not learner outcomes); Emerging evidence indicates generative AI tutors currently struggle to accurately infer student knowledge at the outset of an interaction or as the activity progresses [-W]; K-12 students and teachers both consider AI imperfect at detecting students' knowledge and assigning appropriate tasks, and therefore prefer more human control over AI decisions [~W]
A teacher chooses each learner's materials, path or difficulty by judgement Make each choice from the learner's diagnostic and check results and record the basis beside it; in the weekly review, look for easier materials gathering in one group whose results do not differ, and re-place from results Preservice teachers chose the easy video significantly more often for an ethnic minority student than for an ethnic majority student, all other information being equal [-M]; Student achievement level, not ethnic background, predicted preservice teachers' text-versus-video material choice in Study 2 [~M] (two vignette experiments with preservice teachers, carried to practising teachers; choices of material, not learning outcomes)
Some learners start far behind the rest (lowest prior achievement on the diagnostic) Place them on the continuum where the diagnostic puts them, reaching below grade level if needed, with the same final criterion; track whether they close the gap, not only whether they move; make sure the teacher or tutor goes to them rather than waiting to be asked In mathematics, personalized-learning students in the lowest baseline quintile were the most likely to out-grow their matched comparison students [+W] (matched comparison, mathematics, no effect size); Teachers who made productive adaptations were more likely to consider learning progressions and adapt lessons to below-grade-level content [+W]; who gets reached: Human-AI tutoring is associated with a 36% rise in IXL skill proficiency, but the gap widened in favor of reactively tutored students (difference-in-differences preprint) [~W] (lower starters approached by tutors fell further behind those who asked)
One teacher for a large class, no adaptive software Do not build thirty paths: use the sequenced-sections structure with a short check per section, two or three flexible groups by current section, and a choice board for practice untested proposal; from the folded continuous-progress page ("does not imply a one-to-one teaching ratio")
Facilitator time is scarce (volunteers, a tutor who checks in briefly, a busy shift) Tailor one thing only: pace, through self-checking sections; spend the check-in on the learner who has been held longest at one section, and check afterwards whether those learners closed the gap or only moved Pace and check-in target: untested proposal. Who the facilitator reaches: Human-AI tutoring is associated with a 36% rise in IXL skill proficiency, but the gap widened in favor of reactively tutored students (difference-in-differences preprint) [~W] (a difference-in-differences study; both groups gained, and students tutors approached proactively, who started lower, fell further behind those who asked for help)
Mixed modes (some remote, some in the room) or phone-only Keep the diagnostic and section checks short enough for a phone (five items or fewer); post choices as a list both groups see; review remote learners' progress in the same weekly pass untested proposal
A single session only Personalize only the context of problems and a choice of task; keep one path; record what each learner chose and how they did, and do not claim a lasting gain untested proposal
Learners with low confidence or low literacy Offer two options, read aloud, with a worked sample of each; let learners name interests orally; avoid open-ended choice early untested proposal; the autonomy claim's own discussion warns that premature autonomy can produce anxiety in novices (Self-determination theory holds that autonomy support raises intrinsic motivation, and a meta-analysis finds expected tangible rewards undermine it [~M], carried to this group)
Adults in work settings Use tasks and cases from each learner's own job as the contexts; let them choose which competencies to work on first after a diagnostic, since not all need the same ones Not all students would benefit from improving the same competencies, suggesting an individualized approach to education may be preferable to a one-size-fits-all approach [~W] (an issue brief's analysis of students' competencies and later outcomes; carried to adults in work); contexts: untested
The product or performance is for a real audience Let the learner choose the topic, medium and audience; keep the criteria for the product common and visible untested proposal
The brief names its own routine, categories, criteria or rubric for tailoring (its own check-in, diagnostic, choice menu) Keep it and use its categories; change only what the evidence says matters (diagnose on responses, not styles; ask the learner rather than assign interests; check that the tailoring was enacted) untested proposal
Tailoring needs private or sensitive information (interests, emotions, disability, background) Ask only for what the tailoring uses; the learner decides what to share and who sees it; obtain informed consent and say where it is stored Participants required GSR user-knowledge of learning goals, progress, learning type, emotions, and background, with privacy safeguards and informed consent [~W] (university participants, social-robot design study)
A fixed short time budget (e.g. 30 minutes) Spend 3 minutes on a two-item check, 2 minutes on a choice of context, 20 on practice at the learner's level, 5 on a common unaided check untested proposal
A whole-task performance the brief wants kept whole Personalize the context, topic or means of the whole task, not by splitting it into separately paced parts untested proposal
Stakes are high, or a common standard is set externally Hold the assessment and criterion identical for every path; check that tailored paths still reach every assessed objective before the test untested proposal; from the earlier page's constraint that adaptive routing can narrow exposure

Sixteen of the twenty-three rows cite a claim. Several carry a claim beyond what it tested and say so (the competency row from secondary students to adults, the teacher-judgement row from preservice teachers' vignette choices to practising teachers, the disability-access row from simulated personas, the AI-tutor rows from model outputs, interviews and one scaffolding study, the group-matching row from one gender-of-model experiment and one name-localization trial, the low-confidence row from a claim's discussion, the scarce-facilitator row from paid tutors on one platform to volunteers). Seven rows are untested proposals.

Observation, state and explanation

What can be observed is a learner's response under stated conditions: diagnostic items scored by part, which option or context the learner chose, how long they stayed at a section, their check results, what they said they were interested in, and, for a communication system, which device they selected when free to choose and how many trials a request took. "Is a visual learner", "is interested in football", "is ready to move on" and "prefers the device" are inferred states. An interest a learner names once is a report, not a trait, and it can change.

Response Plausible explanations Observation that could change the interpretation
Engaged with the personalized context but scores no better than on standard problems The context changed interest, not the mathematics; the context added reading load; the learner already knew the skill Give a matched standard-context item; compare reading time and errors; ask whether the context was one they chose.
Always picks the easiest option on a choice board Low confidence; low prior knowledge; the options are not equal in rigour; choosing quickly to finish Check that options share a criterion; offer one item from a harder option with support; ask why they chose it.
Held at one section for two weeks or more A prerequisite gap the diagnostic missed; the check is misaligned with the teaching; disengagement after repeated failure Probe the prerequisite separately; have a teacher go through two failed items with the learner; compare with other learners on the same check.
Says the context "isn't me" The topic was assigned from a group label; the learner's interest changed; the learner does not want mathematics tied to their interests Ask them to name a context; offer a neutral context; record which they keep using.
A learner has been given easier materials or a slower path than peers with similar check results The placement followed an expectation (a group label, an old placement) rather than the work; a need the checks do not show; the learner asked for it Set the learner's check results beside those of peers on the harder materials; offer one item from the harder material with support; ask the learner which they want and why.
A child selects one AAC device but acquires requests faster with another Preference for the device's features rather than its use; novelty; a device suited to one environment and not another Repeat the choice across sessions and settings; record acquisition per device over the same number of trials.

These rows are proposals, untested with learners as diagnostics. A probe is itself an exposure and can change the response it measures, so record each one and keep the explanations that remain open.

Evidence and qualifications

None of the claims below tests personalization as a whole against uniform instruction. Each tests one kind of tailoring, or a limit on it.

Limits and qualifications:

Keep each finding's comparator, population, outcome and horizon with it. Do not rank the undergraduate path experiment beside the school-level SD gains, or read engagement findings as achievement findings.

The evidence by question

Further evidence, not yet read against this model

Claims this page cited that were re-read against the model (on 2026-10-15, and again on 2026-10-19) and left here because they do not bear on its relationship, a step or a situation row beyond naming the topic. On 2026-10-19 the gender-of-model experiment and the two real-time scaffolding claims moved into the model (step 8, step 9, the AI and group-matching rows, the qualifications and the evidence by question). In a second pass the same day, eleven claims the linker had added were read: six moved into the model (the teacher-choice experiments into steps 3 and 9, a new row and a probe; the medical-residents style test into step 8; the AI-STEM school-level meta-analysis into the child row; the lowest-baseline RAND finding into step 4 and a new row; the robot-role interviews into the AI row; the video-preference survey into the engagement evidence), and the five below were left here.

Objective and learner-valued goal

The designer's objective is usually a common capability (the unit's goal, at a stated criterion and horizon) reached by more learners, sooner, with tailoring as the means. Elicit separately what each learner wants from it: to finish faster, to work on things they care about, to avoid being placed on a slower path in front of others, to be able to communicate a request at home, or, for some, not to have their interests tied to a subject they dislike. Agreement and divergence are both observations. A learner who values not being seen as behind may refuse a tailored path that is visible to peers; one who values their own interests may engage with a context and still not learn the structure underneath.

For example, a child learning to request with an AAC system may value being understood by their family at home, while the designer's objective is a number of independent requests in class. The preferred device may serve the first and not the second, or the reverse. Record both, and choose the device with the family's setting in view.

What would revise this model?

The expectation should weaken if experiments that hold content and final assessment constant find personalized paths no better than a fixed sequence on delayed or transfer outcomes, or if the time saving in the one path experiment here does not replicate beyond one topic and one session. If learner-named contexts do not raise interest or achievement more than standard contexts in randomized comparisons, step 6 should be dropped to a motivational option. If preference-matched AAC systems are not acquired or maintained better than non-preferred ones, preference assessment becomes a matter of the child's voice rather than learning. If school-level personalization stays near +0.1 SD even with full implementation, the model's gains belong to the tutoring and diagnosis, not to personalization as a program. Do not protect the model by calling every null result "not personalized enough" after the fact.

Engagement, immediate achievement on an aligned test, time saved, delayed retention and the outcome a learner values are separate claims. The present evidence sets no number of options, no diagnostic length and no pace threshold.

Claims

Examples

Validated

  • Contingent scaffolding improves learning. [+M] — Walkington (2013) personalized algebra word problems to individual learners' out-of-school interests (sports, music, gaming) and found improved performance relative to standard contexts, with the largest gains for struggling learners.
  • Note (2026-10-14): the linked claim page is about contingent scaffolding in tutoring and does not record Walkington (2013); no claim page in the wiki records that study, so the interest-context step above rests on other claims.

Illustrative

Adaptive Learning — Platform-driven personalization in which item difficulty and sequencing adjust continuously to response accuracy and latency. Used at scale in mathematics and language learning; effectiveness depends on mastery gating rather than mere content variety.

ASSISTments — Free web-based math platform (grades 6–12) that routes students through problem sets with step-level hints and immediate feedback, effectively personalizing the support each student receives during practice.

Khan Academy — Mastery-based personalization: learners progress through skill maps at their own pace, with the system recommending review of prerequisite skills when gaps are detected. Widely used in flipped and blended classrooms.

Choice Boards — A teacher-designed personalization pattern: learners select from a curated set of tasks aligned to the same objective, allowing choice of modality or topic while preserving common outcomes. Works best when all options are genuinely equivalent in rigor.

Carnegie Learning MATHia — Intelligent tutoring system that models each learner's knowledge at the skill level and personalizes problem selection and hint delivery, gating advancement on demonstrated mastery.

Duolingo — Language-learning app that personalizes review scheduling (spaced repetition of items the individual learner has struggled with) and adapts session difficulty; a consumer-scale example of adaptive item selection.

Key Sources

  • VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. doi:10.1080/00461520.2011.611369
  • Walkington, C. (2013). Using adaptive learning technologies to personalize instruction to student interests: The impact of relevant contexts on performance and learning outcomes. Journal of Educational Psychology, 105(4), 932–945. doi:10.1037/a0031882
  • Patall, E. A., Cooper, H., & Robinson, J. C. (2008). The effects of choice on intrinsic motivation and related outcomes: A meta-analysis of research findings. Psychological Bulletin, 134(2), 270–300. doi:10.1037/0033-2909.134.2.270
  • Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119. doi:10.1111/j.1539-6053.2009.01038.x
  • Pane, J. F., Steiner, E. D., Baird, M. D., & Hamilton, L. S. (2015). Continued progress: Promising evidence on personalized learning. RAND Corporation. https://www.rand.org/pubs/research_reports/RR1365.html doi:10.7249/rr1365
  • Nam, S., Kim, J., & Sparks, S. (2018). An Overview of Review Studies on Effectiveness of Major AAC Systems for Individuals with Developmental Disabilities Including Autism. The Journal of Special Education Apprenticeship, 7(2). https://eric.ed.gov/?id=EJ1185374
  • Communications: Industrial Arts Curriculum Guide, Middle Schools 6-9 (Tentative). (1974). Louisiana State Department of Education. https://eric.ed.gov/?id=ED120541
  • Bue, T. W. (1979). Teaching and Learning Styles in Higher Education: Match or Mismatch? https://eric.ed.gov/?id=ED172635
  • Hoogerheide, V., Loyens, S. M. M., & van Gog, T. (2016). Learning from video modeling examples: does gender matter? Instructional Science, 44, 69–86. doi:10.1007/s11251-015-9360-y
  • van der Graaf, J., Raković, M., Fan, Y., Lim, L., Singh, S., Bannert, M., Gašević, D., & Molenaar, I. (2023). How to design and evaluate personalized scaffolds for self-regulated learning. Metacognition and Learning. doi:10.1007/s11409-023-09361-y
  • Schlag, S., & Glock, S. (2024). The roles that students' ethnicity and achievement levels play in teachers' choice of learning materials in online teaching: evidence from two experimental studies. Instructional Science. doi:10.1007/s11251-024-09681-9
  • Li, S., Zeng, C., Liu, H., Jia, J., Liang, M., Cha, Y., Lim, C. P., & Wu, X. (2025). A meta-analysis of AI-enabled personalized STEM education in schools. International Journal of STEM Education. doi:10.1186/s40594-025-00566-y
  • Tomlinson, B., Black, R. W., Patterson, D. J., van der Hoek, A., Ferguson, J., & Bietz, M. J. (2026). Students prefer personalized, AI-generated educational videos over non-personalized, human-recorded videos. Scientific Reports, 16, 21804. doi:10.1038/s41598-026-52798-9
  • Vonschallen, S., Oberle, D., Schmiedel, T., & Eyssel, F. (2026). Knowledge-Based Design Requirements for Generative Social Robots in Higher Education. arXiv. doi:10.48550/arxiv.2602.12873
Page metadata
Field Value
type principle
id personalization
aliases assess-child-preference-in-aac-decision-making, continuous-pupil-progress-individualized-structure, four-factor-teaching-compatibility
title Personalization
description Where learners sharing one goal differ in prior knowledge, interests, access needs or preferred means, tailoring support, pace, problem contexts and means to each learner while the goal and criterion stay common is expected to raise engagement and, where tailoring rests on a diagnosis of responses, aligned achievement, though no claim in the wiki tests personalization as a whole against uniform instruction and tailoring to learning styles is contradicted.
status review
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 3 — nam-2018, communications-industrial-arts-curriculum-guide-1974, bue-1979

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