Adaptive Learning

Principle · All principles
Evidence · 69 claims (37 for, 28 mixed, 4 against) · 67 studies (23 causal, 17 review, 8 design, 7 quant-synthesis, 6 associational, 4 theoretical, 2 qualitative), q1–q4 · 11 of 67 report an effect size · 62 claims rest on one study

Conditional relationship

Where learners in one course start from different observed responses on a task whose parts can be checked separately (cumulative procedure, principle or verbal-association goals; mathematics, programming and language mechanics are the usual cases), selecting each learner's next task, hint, feedback or check from that learner's recent responses, instead of moving everyone through one fixed sequence, is expected to raise achievement on an aligned measure (general-achievement, mostly immediate or end-of-course). The relationship is conditional on four things: the responses must diagnose what the learner can and cannot yet do on the target task; the adaptation must change something that matters for learning (the support given or the task chosen, not only its order); support must be reduced as the learner's unassisted performance improves; and the outcome and horizon must be stated, since aligned test scores, standardised achievement, delayed-retention, far-transfer and learning-time move separately.

This page owns the general relationship: individualised selection against a fixed path, whoever or whatever does the selecting. Three converted pages hold narrower models inside it. Scaffolding treats support versus none during an attempt, and its model of contingent support is the within-task form of adaptation. Mastery Learning treats one adaptive rule, a gate with correction and a recheck before advancing. Formative Assessment treats the interpretation of evidence and the contingent action that any adaptive decision depends on. This page does not repeat them. The adaptive learning pattern is a short, unconverted stub.

Only small single studies in the wiki test adaptation itself against the same content without it, and none does so at scale. In the clearest, a randomised experiment with 120 undergraduates, diagnosed personalised paths saved learning time and raised an aligned post-test over a fixed order (Personalized paths reduced learning time and improved post-test scores [+M]); adaptive task difficulty did not help second-graders against an otherwise identical tutor (Adaptive task difficulty did not improve second-graders' posttest [-W]). The strongest evidence below compares intelligent tutoring systems, which bundle adaptation with step-level feedback, practice and a model of the task, with conventional instruction; it shows the bundle can help, not which part of it does.

Default design, while the relationship is untested

The page's earlier guidance, kept as a concrete default a designer can act on and revise. Each step is a design proposal, with its evidence status beside it.

1. Decide what the adaptation is for and how it will be judged. Name the capability, the representative instrument, the criterion and the horizon before choosing a system or rule. Untested here; it follows from the outcome dependence in the tutoring meta-analyses below (gains depended on whether tests were locally developed or standardised).
2. Build a diagnosis the adaptation can rest on. Frequent, fine-grained checks on the parts of the task, using assessment or assess performance. Untested as a step; the earlier page's requirement. The learner-model claim below shows that a model's predictions can be biased where it has little data, so treat each adaptive decision as an inference with error.
3. Provide somewhere to move the learner: a bank or sequence of tasks at varied difficulty (adaptive difficulty), including worked or partly worked examples for low starting responses. The adaptive routing is untested here; starting low responders on an example is supported for novices by one experiment in which sequences opening with a worked example beat problem-first sequences on an immediate test, with lower reported effort (Example-problem sequences reduce cognitive load [+M]; secondary-school electrical troubleshooting, n = 96 analysed; delayed and far-transfer outcomes not tested).
4. Adapt the support, not only the difficulty or order. Give hints and task- and process-level feedback tied to what the response showed, raising support after an error and lowering it after success. Supported in one-to-one tutoring by the contingent-scaffolding claim. What the feedback should carry is supported, from feedback research in general rather than adaptive systems, by two meta-analyses in which high-information feedback on the task and process outperformed reinforcement and self-directed feedback (Feedback is most effective at task and process levels [+S]; both support the levels only indirectly). In AI tutoring logs, the association between asking another question and recovering from an impasse weakened with each further failed turn, while addressing the learner's error became more useful (Questioning benefit decays with impasse depth [+W]; Repeating a scripted question lowers recovery [+W]; one associational study). The earlier page's statement that difficulty-only adaptation gives weaker gains is not tested by any claim here.
5. Gate progression on an explicit criterion with correction and a recheck (adaptive mastery learning), as modelled in Mastery Learning. Evidence is on that page.
6. Withdraw support as unassisted performance grows, checking on unsupported items before removing it. Bounded support from the expertise-reversal claim below (a review; abstract not confirmed), and from fading worked examples into independent problems, which favoured near transfer over static example-problem arrangements in three experiments (Fading support promotes the transfer of responsibility [+S]; near transfer only, no fading rate established).
7. Do not adapt to self-reported learning styles. Against: the learning-styles claim below.
8. Watch for narrowing and misrouting. Check transfer and conceptual items the system does not optimise for, and review learners routed to remediation repeatedly or advanced quickly on sparse evidence; if the system optimises a reward of its own, set it on unsupported learning, not engagement or in-system correctness (RL reward misalignment failure modes [~W], an analytical chapter). Untested; from the earlier page's constraints.
9. Check that the adaptation is actually enacted (usage, teacher use of the data) before judging its effect. Bounded support from the personalised-learning fidelity claim below.

How far the neighbouring claims carry: the tutoring meta-analyses are about whole systems against non-individualised instruction, not about any one step above; the contingent-scaffolding evidence is mostly one-to-one human tutoring; the fidelity claim is a matched-comparison study of school-wide personalised learning, not of an adaptive rule.

Fitting the design to a situation

Ask for four facts when the brief omits them: what the system will adapt (the support after a response, or only difficulty and order), what evidence it adapts on (responses to items, or self-reported preferences), which outcome and horizon count (the system's own items, an aligned test, a standardised test), and who will look at the learner data. The strongest evidence compares whole tutoring systems with conventional instruction; only a few studies compare adaptation with the same content in a fixed order.

If the brief says…	Then change…	Basis
The alternative is conventional, non-individualised instruction or a textbook	Use a system that selects tasks and step-level feedback from each learner's responses; expect a gain on aligned tests, smaller on standardised ones, and none over a human tutor or a small group	Intelligent tutoring systems improve achievement over non-individualised instruction [+S]
The plan adapts only task difficulty, for young learners (early primary)	Add adaptive support (hints, worked steps, feedback tied to the error) rather than relying on difficulty alone; one trial found difficulty-only adaptation no better than a fixed version, and its mastery gate held learners back	Adaptive task difficulty did not improve second-graders' posttest [-M]; The non-adaptive condition reached higher mastery [-M]
University learners, a cumulative topic with checkable parts	Diagnose the parts with frequent short items and order each learner's path from the diagnosis; expect time saved as well as a modest post-test gain	Personalized paths reduced learning time and improved post-test scores [+M]; Adaptive scaffolding policies improved posttest in a logic tutor [+M]
A tutor (human or AI) responds to learners who are stuck	Change the support as the impasse persists: after a failed question, address the learner's error rather than repeating the question	Questioning benefit decays with impasse depth [+W]; Repeating a scripted question lowers recovery [+W]
An AI tutor is used, and the platform holds mastery and practice history	Pass the learner's mastery estimates for the skill and its prerequisites to the tutor, with an instruction to review weak prerequisites	Student mastery context improved tutor metrics [+M]; Generative AI tutors struggle to infer student knowledge [~M]
Learners improve quickly on a skill	Withdraw worked steps and hints as unassisted performance improves, checking unsupported items before each removal	Guidance that helps novices can become redundant as expertise grows [~M]; Fading support promotes the transfer of responsibility [+S] (near transfer)
Learners choose their own next tasks inside a mastery system	Constrain choice where a learner's selections are inefficient (for example, risk-averse overpractice of mastered skills) and leave efficient selectors free, rather than constraining everyone	Constraints selectively affect inefficient strategies [~W] (simulated learners, not tested with people)
The brief proposes adapting to learning styles or stated preferences	Adapt on responses to items and on prior knowledge instead	Learning-styles matching does not improve learning [-S]; in one series of media experiments a voice-plus-diagram format lifted sensing-preference learners above intuitive-preference learners given text, so format, not the match, drove the difference (Intuitive learners outperform sensing learners [-W])
Routing depends on a knowledge-tracing or learner model	Treat estimates at the extremes (very low or very high predicted success) as least reliable: add a teacher check or an unassisted probe before repeated remediation or a fast skip	Learner models are miscalibrated outside the data interval [~M]
A vendor reports the model's predictive accuracy (AUC)	Ask whether attempts were ordered by time in the evaluation, and how the model is calibrated, not only its AUC; leakage inflates reported accuracy	Violating temporal order inflates knowledge-tracing performance [~M]; A transformer model won EDM Cup 2023 with an AUC of 0.79 [~M]
A school or adult programme adopts a platform	Plan for teachers to read and act on the diagnostic reports, and record use; judge the effect only where the adaptation was used	Personalized learning effects vary with fidelity [~M]; Adult education programmes used learner data minimally [~M]
The system optimises its own reward (a reinforcement-learning policy)	Set the reward on learning measured without support, not on engagement or in-system correctness, and watch for learners who depend on scaffolding	RL reward misalignment failure modes [~W]

Observation, state and explanation

An adaptive system observes responses to particular items under stated conditions (correctness, hint requests, time, attempts, which help was shown) and turns them into an inferred state: "has mastered this skill", "probability correct is 0.8", "ready for harder items". The inferred state is a model output, not an observation, and it inherits the model's errors. "Struggling" and "advanced" are labels for a learner's responses on this task, not traits. Record the items, the help given and the model's estimate side by side, so a routing decision can be checked against what was actually seen.

Response	Plausible explanations	Observation that could change the interpretation
Correct on the system's items, fails a teacher's or standardised test of the same topic	System items narrower than the target; capability tied to the system's format or hints; forgetting between practice and test	Give an item on the same relation in a different format and without hints, soon after practice and again later; compare which conditions the system's items omitted.
Many hint requests, then correct answers	Learning from the hints; using hints to reach the answer without the reasoning; low confidence rather than low knowledge	Present a comparable item with hints withheld and ask for the first step; ask the learner why they asked for help.
Routed to remediation repeatedly on one skill	A prerequisite gap the system does not model; a skill estimate biased where the learner's data are sparse or extreme; items misaligned with the instruction; disengagement after repeated failure	Probe the prerequisite separately; have a teacher check two or three of the failed items with the learner; compare with other learners' results on the same items.
Advances quickly through items	Prior mastery of the target; guessing or gaming on easy items; criterion set too low	Give a few harder or transfer items unassisted; look at response times and error patterns on the items passed.
Gains in a school using the platform	The adaptation; extra practice time; the teacher's use of the data; selection of which classes used it	Compare classes with similar starting scores and similar time on task; record how fully the platform and its reports were used.

These rows are proposals, untested with learners. A probe can shift confidence among explanations; one contrast does not identify a cause, and the probe is itself an exposure that can change the response. Record what remains unresolved rather than letting the system's estimate settle it.

Evidence and qualifications

• Intelligent tutoring systems improve achievement over non-individualised instruction [+S]: two meta-analyses (synthesis-experimental, heterogeneous-synthesis). Across 107 effect sizes (14,321 participants), intelligent tutoring systems produced higher achievement than teacher-led large-group instruction (g = 0.42), non-tutoring computer-based instruction (g = 0.57) and textbooks or workbooks (g = 0.35), but no advantage over individual human tutoring (g = -0.11) or small-group instruction (g = 0.05). Across 50 controlled evaluations, the median effect over conventional instruction was 0.66 SD, depending strongly on whether tests were locally developed or standardised, and small where controls were nonconventional or implementations flawed. Both entries pass the judge, on abstracts. It supports individualised tutoring over non-individualised instruction on aligned achievement; it does not isolate adaptation from the rest of the system, does not show an advantage over adaptation by people, and reports no common horizon.
• Contingent versus fixed support [+M]: in one-to-one tutoring of long division with fourth and fifth graders (n = 8 per condition), fully contingent support beat fixed, moderate, partly contingent and no support on immediate and one-month follow-up tests; a within-subjects tutoring study reports similar immediate outcomes but better transfer for interactive, contingent tutoring; a synthesis of dynamic assessment ranks explicit strategy training above contingent scaffolding. This is the closest test here of adapting support to responses against fixed support, and it is small and mostly one-to-one; classroom and software contingency are not shown. Not settled: the entries could not be confirmed from abstracts.
• Guidance and expertise [~M]: a review of cognitive load studies reports that supports such as worked examples and integrated explanations help novices on a task and can become redundant or depress performance as prior knowledge grows. It qualifies the relationship: adaptation that keeps support in place as performance improves can work against it. No effect size; not settled on its abstract.
• Personalised learning and implementation [~M]: in about 40 schools implementing personalised-learning practices (adaptive software among them), matched against virtual comparison groups (controlled-nonrandom), one-year effects were small (+0.09 SD mathematics, significant; +0.07 SD reading, not significant) and larger where implementation was reported as fuller or longer; the authors call that link suggestive, resting on self-report and a small district subsample. It qualifies the relationship: an adopted platform is not an enacted adaptation, and school-level averages are far below the tutoring syntheses' figures. Not yet checked against its source.
• Learning-styles matching [-S]: a systematic review found almost no studies with the crossover design needed to test matching instruction to learning style, and judged the one candidate unconvincing; a randomised study with fifth graders found no style-by-modality interaction. It bears against adapting on self-reported style; it says nothing about adapting on responses or prior knowledge. Not settled on abstracts.

• Learner-model calibration [~W]: one analysis of two knowledge-tracing models on tutoring datasets reports that both were calibrated only in the range containing most of the data, overestimating learners when the probability of a correct answer was low and underestimating them when it was high. It limits the diagnostic step: a system's estimate for a learner at the extremes is least trustworthy where routing matters most. Associational, one study, no outcome for learners; not yet checked against its source.

• Tutor moves at an impasse [+W]: in logs of guided AI tutoring (6,163 impasse turns), the association of scripted and follow-up questions with recovery declined with each further impasse turn while addressing the learner's incorrect answer rose (AORs 0.78, 0.83 and 1.14 per turn); repeating a failed scripted question was followed by recovery in 28.1% of cases against 39.8% after switching to the error (Repeating a scripted question lowers recovery [+W]). It supports adapting support to the persistence of a difficulty within a task; it is associational, one study, measures recovery within the session rather than learning, and does not test adaptive against fixed sequencing.

• What the adaptation changes [+W]: one 1972 report on adaptive training of a single-axis manual tracking task (eight groups of eight) found the gain-effective time constant product slightly better than forcing-function amplitude as the adaptive variable, and system compensation, aiding and quickening unsatisfactory [~W]. The same report found manual adaptation slightly superior to automatic adaptation [~W], which the authors say could reflect different performance-measurement levels. It bears on the second condition (what the adaptation changes matters) in perceptual-motor training, not in academic content; no effect sizes.
• Speed of acquisition and retention [~W]: a review of skill-learning research reports that slower adaptation processes were more robust over time and that slower learners in sequence learning retained more. It qualifies reading time saved to criterion as a gain: a design that moves learners on as fast as possible should check delayed-retention. Motor and sequence learning, cited second-hand, no effect size.

Keep the comparator (conventional, non-tutoring software, human tutor), the population band, the outcome instrument and its alignment, and the horizon with each finding. Do not set the meta-analytic g values beside the school-level SD gains as if they measured the same thing.

The evidence by question

Each question below links a general claim that gathers the studies behind it into one table, one row per distinct study, with each study's design and effect as recorded and no averaged verdict (eval/synthesis/adaptive-learning.json, scripts/synthesize_claim.py). The count is of distinct studies across the claim and the narrower claims that bear on it.

• Does adaptive instruction raise achievement over non-individualised instruction or the same content in a fixed sequence? Intelligent tutoring systems improve achievement over non-individualised instruction [+S]: 10 distinct studies.
• What should adapt: the support given after a response, or only difficulty and order? Contingent scaffolding improves learning more than fixed or absent support. [+S]: 20 distinct studies.
• When should support be withdrawn as the learner improves? Instructional guidance that helps novices can become redundant or counterproductive as expertise grows. [~S]: 13 distinct studies.
• What should the adaptation be based on: responses, or learner styles and preferences? Learning Styles Matching Does Not Improve Learning [-S]: 4 distinct studies.
• Can the system's diagnosis of the learner be trusted? Current best learner performance models are severely biased outside the interval containing most of the data, hindering downstream adaptive policies and open learner models [~M]: 7 distinct studies.
• What makes an adaptive design fail in practice? Personalized Learning Effects Vary With Fidelity [~M]: 4 distinct studies.

Further evidence, not yet read against this model

Re-read 2026-10-15: the claims below were read against the model above and do not bear on its relationship, steps or situation rows; each label says how far they reach. The others once listed here are now worked into the model.

• Integrating cognitive scores with non-cognitive learner attributes enables the system to dynamically adapt AI-generated explanations in tone, depth and instructional strategy to individual learner performance [+W] — read 2026-10-15, does not bear on the relationship: a two-case demonstration that a system's explanations differ with a performance score, with no learner outcome measured.
• Diminishing step sizes in stochastic approximation induce implicit averaging that eliminates noise effects in the long run [+W] — read 2026-10-15, bears only on the estimation machinery behind an adaptive rule (a mathematical result about noisy sequential estimation), not on learners.
• Iterate averaging with step sizes decreasing slower than O(1/n) converges at an optimal rate [+W] — read 2026-10-15, as above: a convergence result for sequential estimation, no learners.
• Propdec Colinear With Student Intercept [+M] — read 2026-10-15, bears only on how a learner model represents individual differences (model fit across datasets); no routing decision or learner outcome, so it does not test the diagnostic step.
• Kt Encoders Entity Isolated Deterministic Limits [+W]
• Adaptive G Ukt Competitive Performance Sparse Regimes [+M]
• Identifying learners' difficulties with visual displays is an important basis for selecting appropriate support procedures [+M] — attached 2026-10-15 from Renkl et al. (2017), which proposed "Choose support procedures for visual displays based on knowledge of the difficulties learners have with them".
• Intelligent tutoring systems with dynamic student modeling increase learning gains over regular instruction [+W] — attached 2026-10-22 from Leonard Tetzlaff et al. (2021), which proposed "Measure relevant learner characteristics repeatedly during learning and adapt instruction systematically"; tests this page's relationship.
• Degree of adaptivity moderates the adaptive teaching effect: only highly adaptive units outperform control, and boost self-efficacy [+M] — attached 2026-10-22 from Sibley et al. (2025), which proposed "Implement adaptive teaching with high fidelity: apply adaptive elements constantly, since benefits emerge only in highly adaptive units"; tests this page's relationship.
• Subject domain moderates the adaptive teaching effect: benefits in languages, none in STEM or humanities [~M] — attached 2026-10-22 from Sibley et al. (2025), which proposed "Implement adaptive teaching with high fidelity: apply adaptive elements constantly, since benefits emerge only in highly adaptive units"; tests this page's relationship.
• Individual records revealed contrasting specific disabilities within the same pupil across arithmetic operations [+W] — attached 2026-10-24 from Buford Johnson (1918), which proposed "Instruction should be devised from elaborated testing built on analysis of individual performance".

Objective and learner-valued goal

The designer's objective in an adaptive design is usually a criterion on each skill in a map, reached in less time than a fixed path would take. What the learner values may differ: finishing the assigned work quickly, understanding why a method works, preparing for a particular examination, or not being held on one skill while classmates move on. Ask what the learner wants and why, record agreement or divergence as an observation, and do not infer it from time logged in the system.

For example, a secondary-school learner using an adaptive mathematics platform may value being ready for an end-of-year standardised examination, while the platform's objective is mastery of each skill on its own items. The tutoring synthesis above reports that gains depended on whether tests were locally developed or standardised, so the two aims may not move together. A design that serves both states which instrument and horizon count, adds unassisted examination-style items to the platform's checks, and tells the learner how the platform's mastery estimate relates to the examination.

What would revise this model?

The expectation should weaken if, with comparable learners and content, individualised selection does not outperform the same content in a fixed sequence on independent outcomes, or if tutoring-system gains disappear on standardised measures and at a delay. If difficulty-only adaptation does as well as adaptation of support, the claim that support is the active part should be revised. If routing decisions based on a learner model are no better than a teacher's or the learner's own choices, the diagnostic step needs rethinking. If time saved, persistence or the learner's own goals move against achievement, the model is incomplete without them. Do not protect it by relabelling every failure as low fidelity or poor engagement.

Within-system performance, aligned achievement, standardised achievement, delayed retention, transfer and time to criterion are separate claims. The present evidence does not establish which adaptation target matters most, an optimal difficulty band or mastery threshold, or a forecast for an individual learner.

Claims

• Adaptive learning systems report academic performance gains of 15–25% and engagement increases of up to 40% [~W]
• Adaptive scaffolding policies (BKT and DRL) significantly improve posttest performance over a non-adaptive control in a logic ITS [+W]
• Adaptive fading with periodic reinforcement yields higher independence than fixed-high or linear-decay scaffolding in simulation [~W]
• The survey reports, citing Desmarais and Baker, that students using the BKT-sequence recommendation algorithm solved more difficult exercises, obtained higher performance and spent more time in the system than students using the traditional approach. [+W]
• Case study shows tutor scaffolding lowers estimated task difficulty after errors while student ability rises steadily [~W]
• GenAI can provide personalised feedback in group work, but the depth of personalisation is contested [~W]
• The system's generated practice trajectories align with learner skill levels at a peak cosine similarity of 0.962 ± 0.014 at the primary difficulty level [~W]
• LMS process indicators documented adaptive engagement during the eight-week intervention, with most experimental students reaching mastery by unit end [+W]
• Learner engagement was strengthened when the AI tutor adapted to learners' pace and prior knowledge [+W]
• One-shot path generation raises Validity but lowers Adaptivity and Efficiency relative to step-by-step interaction [~W]
• Students in the non-adaptive condition reached significantly higher average mastery probabilities than students in the adaptive condition [-W]
• 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]
• In a randomized experiment, personalized paths reduced learning time by 22.0% and improved post-test scores versus a fixed sequence [+M]
• Personalized paths reduced self-reported cognitive load on all six adapted NASA-TLX dimensions with large effect sizes [+W]
• Tutors showed significant pre-to-post gains in both conceptual knowledge and procedural skill after teaching with adaptive follow-up questioning [~W]
• Providing the tutor with student mastery and practice-history context improved engagement and next-item correctness [+W]
• Adaptive task difficulty in an ITS did not improve second-graders' mathematics posttest performance compared with an identical non-adaptive ITS [-W]
• Reinforcement learning in adaptive educational systems faces sample inefficiency and reward-design instability, motivating hybrid RL-ML approaches [~W]
• An adaptive AI-based STEM instructional program produced statistically significant differences favoring the experimental group across all deep learning dimensions in sixth-grade science [+W]
• Meta-analytic evidence reviewed reports medium-to-large positive effects of AI-enabled adaptive learning on cognitive outcomes (g = 0.50–0.70) [+M]
• Adaptive versus non-adaptive ITS condition showed no significant differences in mental effort, subjective difficulty, or enjoyment [~W]
• Reviewed literature identifies personalisation, adaptive support, rapid feedback, and teacher-facing automation as recurring pedagogical opportunities of AI in mathematics education [+W]
• The EEG-adaptive exercise condition produced significantly higher cognitive engagement during training than the unrestricted chatbot condition (p = .018), and was the only condition above the 0.5 baseline [+W]
• Effectiveness of AI-driven personalized learning is context-dependent, with CS and Mathematics students showing contrasting engagement and performance trajectories [~W]
• A multi-site randomized controlled trial found Woot Math's adaptivity contributed significant effects in learning and retention [+W]
• A technical review reports that one reinforcement-learning adaptive platform raised average course grades by 12%, task completion by 20% and cut dropout by 10% against static curricula (uncontrolled case report) [+W]
• A stochastic (exploration) router yields a higher exercise conversion rate than the static baseline, while the greedy router is comparable [+M]
• Adult education programs used product-generated learner data minimally, mostly tracking usage time rather than diagnosing struggles [-W]
• The DRL policy produced a significantly different scaffolding distribution than Control and BKT, favoring Guided examples (60%) and avoiding Buggy examples (4%) [+W]
• In a classroom study of the G4L system, the Bayesian knowledge propagation algorithm demonstrated the highest knowledge gains, and log-data analysis revealed significant differences in self-regulated learning patterns between learners [~W]
• Individualized Knowledge Tracing parameters fit student data significantly better and change mastery-learning practice decisions for many students [~W]
• Exploratory ANCOVA controlling for mastery probability favored the adaptive condition, but is not interpretable as a causal treatment effect [~W]
• A multi-skill prior strategy yielded lower prediction error in 33 of 42 problem sets [~W]
• Adaptive scaffolding shows descriptively higher effects than adaptable scaffolding in simulation-based learning, though the difference is not statistically significant [+W]
• Adaptable task progression is descriptively more beneficial than fixed task progression in simulation-based learning [+W]
• Technology-supported adaptive teaching units showed significant learning gains, larger for low-prior-knowledge students and high implementation fidelity [+W]
• System-controlled adaptive instruction does not necessarily yield higher transfer than nonadaptive static approaches [~W]
• ITS augmentation: meta-analyses find medium to large learning benefits of ITS over unsupported learning and small to medium gains over less adaptive computer-assisted learning [+M]

Related Principles

• Cognitive Load Management — adaptation is the primary mechanism for keeping intrinsic load matched to learner expertise over time
• Assessment for Learning — supplies the continuous diagnostic evidence that any adaptive decision depends on
• Competency-Based Learning & Assessment — provides the mastery criteria that gate progression in adaptive designs
• Active Learning — adaptive systems still require learners to do generative work; adaptation of difficulty does not replace engagement
• Mastery Learning
• Knowledge Tracing Learner Modeling Task
• Knowledge Tracing Driven Exercise Recommendation
• Adaptive Difficulty
• Engagement

Examples

• Four-layer system architecture for intelligent oral diagnosis and adaptive training
• Mask the reinforcement learning policy's action space to a zone-of-proximal-development difficulty band (success probability 0.4–0.8)
• Use shaped-spectrum filtered error for rapid adaptive logic response
• Four ecosystem recommendations: experiment with currency forms, protect time for inquiry, include everyone in team learning, and invest in adaptive platforms
• Use early student-model predictions to identify at-risk students and adapt pedagogical strategy during tutoring
• Use in-game scaffolds with automatic feedback displays in math game tasks, monitoring scaffold use against mastery data
• Use AI to support autonomy-supportive task adaptation in physical education rather than algorithmic exercise prescription
• cc-self-train adaptive learning system: a three-layer hook-based engagement observation and scaffolding adjustment mechanism
• Use traced error causes as personalized intervention recommendations in online learning systems
• Select iteration coefficients according to the proved convergence condition when implementing stochastic approximation
• Iteration-coefficient sequences as the controllable design element of stochastic approximation
• Design AI learning systems for inspectable, contestable use: provenance, represented uncertainty, user control and epistemic plurality
• Self-improving mechanism: learning strategy-selection criteria from student model states by concept learning from examples
• For young learners, adapt instructional support level rather than only task difficulty, and investigate adaptive features in isolation
• Use strategy-based evidence models to trigger adaptive scaffolding toward more productive problem-solving strategies

Validated

• ASSISTments — Free web-based math platform (grades 6–12) that adapts problem selection and hint delivery based on item-level responses. Randomized studies across Maine schools showed significant homework-related learning gains over business-as-usual conditions (Roschelle et al., 2016, AERJ).
• Carnegie Learning MATHia — Cognitive-tutor-based adaptive math system using a cognitive model of learner knowledge to select problems and tailor step-level hints. A large RAND study (Pane et al., 2014) found roughly doubled learning-growth effects in second-year algebra relative to conventional instruction.
• Khan Academy — Mastery-based practice in mathematics that adapts task assignment to demonstrated skill levels, with mastery gates before progression.

Illustrative

• Adaptive Difficulty — The core element: dynamically raising or lowering task difficulty based on performance signals such as accuracy, latency, and error patterns.
• Adaptive Mastery Learning — Combines mastery criteria with adaptive routing so learners who fail an assessment are routed to remediation rather than the next unit.
• Duolingo — Language-learning app that adapts item scheduling using a spaced-repetition and learner-error model, reinserting items the learner is predicted to forget.

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
• Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of intelligent tutoring systems: A meta-analytic review. Review of Educational Research, 86(1), 42–78. doi:10.3102/0034654315581420
• Ma, W., Adesope, O. O., Nesbit, J. C., & Liu, Q. (2014). Intelligent tutoring systems and learning outcomes: A meta-analysis. Journal of Educational Psychology, 106(4), 901–918. doi:10.1037/a0037123
• Corbett, A. T. (2001). Cognitive computer tutors: Solving the two-sigma problem. User Modeling 2001, 137–147. doi:10.1007/3-540-44566-8_14
• Pane, J. F., Steiner, E. D., Baird, M. D., Hamilton, L. S., & Pane, J. D. (2017). Informing progress: Insights on personalized learning implementation and effects. RAND Corporation. https://www.rand.org/pubs/research_reports/RR2042.html doi:10.7249/rr2042
• Pane, J. F., Steiner, E. D., Baird, M. D., & Hamilton, L. S. (2015). Continued progress: Promising evidence on personalized learning. RAND Corporation. doi:10.7249/rr1365
• Shuanghong Shen, Qi Liu, Zhenya Huang, Yonghe Zheng, Minghao Yin, Minjuan Wang, and Enhong Chen. (2021). A Survey of Knowledge Tracing: Models, Variants, and Applications. https://arxiv.org/abs/2105.15106
• Norman, D. A.; And Others. (1972). Adaptive Training of Manual Control: 1. Comparison of Three Adaptive Variables and Two Logic Schemes. Life Sciences, Inc. https://eric.ed.gov/?id=ED059585 doi:10.1037/e426822004-001
• Wolff, Hans. (1970). On Stochastic Approximation. https://eric.ed.gov/?id=ED051258
• Philip I. Pavlik, Jr., Luke G. Eglington, and Leigh M. Harrell-Williams. (2021). Logistic Knowledge Tracing: A Constrained Framework for Learner Modeling. IEEE Transactions on Learning Technologies. https://doi.org/10.1109/TLT.2021.3128569
• Jia Nan, Su Weitao, Xian Junrui, Zou Shijia, Xia Yixue. (2026). Adaptive G-UKT: a unified probabilistic framework for knowledge tracing via adaptive graph topology learning and uncertainty-aware Gaussian embeddings. Scientific Reports. https://doi.org/10.1038/s41598-026-50711-y

Page metadata

Field	Value
type	principle
id	adaptive-learning
aliases	adaptive-learning-consistent-with-cognitive-structure, gain-then-time-constant-adaptive-progression, multi-objective-exercise-recommendation-objectives, sequential-noisy-estimation-principle, avoid-student-intercepts-for-transfer, uncertainty-aware-knowledge-states-for-adaptive-decisions
title	Adaptive Learning
description	When learners on a cumulative, decomposable task start from different observed responses, choosing each next task, hint or check from the learner's own recent responses (rather than one fixed path) may improve aligned achievement, conditional on how well the responses diagnose the learner, what the adaptation changes, and whether support is withdrawn as performance grows.
status	review
moved_from	pattern/adaptive-learning
canonical	True
generated	by claude/unspecified on 2026-10-15
trust	unverified — no one has confirmed this page's content
sources	5 — shuanghong-shen-2021, norman-1972, wolff-1970, philip-i-pavlik-2021, jia-nan-2026

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