Skip to content

Cognitive Load Theory

Principle · All principles
Evidence · 65 claims (26 for, 38 mixed, 1 against) · 62 studies (28 causal, 14 review, 13 theoretical, 6 quant-synthesis, 1 qualitative), q1–q4 · 13 of 62 report an effect size · 54 claims rest on one study

Conditional relationship

For a learner who is a novice on the particular task (not in the wider domain) and who meets material high in element interactivity (elements that must be processed together), presentations and sequences that remove avoidable processing (integrated rather than split sources, segmented transient media, isolated interacting elements before the whole, full rather than completion examples) are expected to improve immediate performance, near transfer or instructional efficiency, compared with the unreduced version of the same material. The relationship weakens, disappears or may reverse as the learner's task-specific expertise grows, and it is expected to matter little for low-element-interactivity material. Which outcome changes (performance, effort, efficiency) and at what horizon (immediate, delayed) differs between the studies below and must be stated for each design. Working-memory overload and schema construction are the explanatory hypotheses of Cognitive Load Theory and the expertise reversal effect; they are not observations. The 4C/ID pattern is a reusable whole-task policy that applies this relationship; its response-dependent branches remain proposals.

Fitting the design to a situation

Ask for three facts when the brief omits them: how much do these learners already know about this particular task?, how many elements must they handle together (simple or complex material)?, and how is it presented: text and diagrams, narration, video, and who sets the pace?

If the brief says… Then change… Basis
Text and diagrams refer to each other (labelled figures, worked solutions with a diagram) Place each label or explanation on or next to the part it describes; present narration at the same time as the animation it explains. Where the text only restates what the diagram already shows, do not expect removing or keeping it to matter much Reducing split attention improves learning [+S]; Integrating text and diagrams beats separating them [+M]; text repeating a diagram's frequencies made no detectable difference [~M] (198 undergraduates, a print probability lesson; not an equivalence test; see Redundancy)
Pictures or an animation are explained by printed words, and the lesson runs at the system's pace Speak the explanation instead of printing it beside the picture; keep printed words to labels; see Modality Animation or picture with voice beat picture with text [+M] (36 adults, a 15-minute immunology lesson, p = .004, no effect size printed)
The material is complex, with many interacting elements Apply the load-reducing design principles first here; their effect was largest for complex material Reducing extraneous load improved learning, more for complex material [+M]; Integration helps more for high element interactivity [+M]
The material is simple, or learners control the pace (a website, a document) Expect a smaller gain from redesign; reported effects were smaller for simple and for learner-paced media Design principle effects moderated by pacing and complexity [~M]
Narration plus the same words on screen, alongside pictures or animation Drop the verbatim on-screen text; with images or animation, adding it gave no benefit. Removing text that is already there is a small change, not a rescue With images or animation, verbal redundancy does not help [~M]; adding written text to narration made no significant difference for middle-school students [~M] (abstract and concrete animation; a quasi-experiment, source without a DOI)
Learners are second-language or less fluent readers, and the material is text without pictures Keep spoken and written text together (narration with captions); the benefit was larger for these readers Spoken–written advantages are larger for L2 and nonfluent readers [+M]
The instruction is in a language the learners are still learning Give explanations in the learners' first language or bilingually where possible Foreign-language AI explanations impose a double burden [~W]
Second-language readers of different proficiency share one set of reading materials Give each proficiency level its own version: glosses and sentence unpacking for the least proficient, different support for the others, checked by a comprehension test per level Differentiated materials beat one adapted version within each proficiency level [+M] (one EAP study: d = 0.60 low, 0.16 intermediate, 1.40 high); lower rated load for low-proficiency readers with differentiated materials [~M] (descriptive only, measured in the same session as the outcome)
Learners are experienced on this task Remove explanatory text from diagrams and other guidance that repeats what they know; check with an unsupported task Guidance that helps novices can become redundant as expertise grows [~M]; Split-attention benefit reverses into redundancy for experienced learners [~M]
Learners are novices on a multi-step procedure Open with a full worked example, not a problem: examples only or example–problem pairs, then problems; use full worked examples rather than completion problems at first Example-first sequences beat problem-first ones for novices on an immediate test, with lower rated effort [+M]; Lower-prior-knowledge learners did better with full than completion examples [+M]
Support can be adjusted to each learner's last response (one-to-one tutoring, a small group, an adaptive system) Raise support after a failed step and lower it after a success, instead of fixing one level for everyone; fade full examples as independent success holds Contingent support beat fixed or absent support, mostly in one-to-one tutoring [+M]; Worked-example guidance becomes less effective as expertise grows [~M]
A complex whole task that novices cannot yet coordinate Teach the interacting parts separately first, then the whole task Isolated parts first helped novices on complex tasks [+M]
Learners have an AI assistant or tutor Let it remove access barriers (finding, formatting, translating material) but keep the interacting elements the task is meant to teach for the learner to process; teach learners to use it as a partner that questions and checks rather than an answer source; check unaided performance, and do not expect rated load to fall One trial in the wiki: structured training to use generative AI as a collaborative partner raised electromagnetism posttest scores over problem-solving training [+M] (sophomore STEM students, d = 0.81) while total rated load did not differ and intrinsic load was marginally higher [~M] (d = 0.17; d = 0.37, p = .079). Otherwise conceptual arguments: AI lowers access complexity, not the task's cognitive demand [~W]; Outsourcing productive struggle to AI risks cognitive debt [~W]; AI as scaffolding that lowers the threshold for understanding [~W]; Intrinsic and extraneous load differ by whether interactivity is essential to the task [~W]
Learners work in groups on a shared complex task Keep the amount of information each group must pool and coordinate low at first; prefer groups that have worked together when it is dense; check retention after a delay, not only the next day. Group size and structure are on Cognitive Load Management Low information density lowered rated effort and raised efficiency during group learning [+M] (one 2×2 experiment); its retention advantage had faded by the delayed test [~M]
The redesign would remove work learners are meant to learn from (comparing examples, relating two representations, choosing a path through linked material) Before removing it, ask whether that processing is the target; if it is, keep it and reduce load elsewhere (pre-teach parts, integrate sources), then compare learning, not ratings Reducing extraneous load can also remove affordances for germane processes [~W] (de Jong's review, citing a hypertext example; an argument, untested in the wiki)
Success will be judged by learners' effort or load ratings Measure performance and transfer as well; load ratings moved little even where design changes improved learning, and moved without any change in learning in a combinatorics simulation and a case-comparison study; an added support can raise ratings, so a higher rating alone does not show the support failed Design interventions yield small improvements in load measures [~M]; Tree diagrams raised rated load [~M], yet format made no difference to conceptual or situational knowledge [~M] and adding equations lowered load without better outcomes [~M]; principles before case comparison raised rated germane load and lowered extraneous load by large amounts [~M] while timing made no difference to knowledge [~M] (student teachers); drawing support raised rated difficulty [-M]

Observation, state and explanation

Record a response under stated conditions: the task, its representation, its element interactivity as judged for this learner, the support present (integrated text, segments, examples, hints), time and the response mode. "Overloaded" and "novice" are inferred states, not observations. Accuracy, time, a rated mental effort and confidence are observations; none directly measures working-memory load or a schema, and a low effort rating can also mean disengagement. Expertise is task-specific: a learner fluent on one task class may be a novice on the next, so a rating from an earlier topic does not settle it. Load ratings also travel with motivation: in one algebra study, self-reported extraneous load correlated negatively with germane load (r = −.26) and with expected probability of success (r = −.19), and germane load correlated positively with interest (r = .26), all self-report and correlational (extraneous load correlates negatively with germane load and expected success [~M]; germane load correlates with interest [~M]). Read a load rating alongside interest and expected success, not as a pure measure of the material. A review makes the same point from theory: cognitive load theory has little to say about motivation [~W], so the effort a learner chooses to invest (in retrieval or elaboration, say) is not the same thing as the effort the material's complexity imposes, and a design that raises the first is not explained by load reduction.

Response Plausible explanations Observation that could change the interpretation
Fails a complex task but succeeds on each component in isolation Interacting elements exceed what the learner can coordinate; a missing integrating relation; a presentation that splits mutually referring sources Present the same whole task with the sources integrated, then with one interacting element pre-taught; record which change restores performance and every support given.
High rated effort and slow progress with full support Material genuinely high in element interactivity for this learner; redundant support the learner must reconcile with what they already know; the support itself adds processing (a drawing aid raised rated difficulty against none, drawing support raised perceived difficulty [-M]); unfamiliar representation or language Offer the same task with the redundant support removed (e.g. diagram without duplicating text) and compare accuracy and effort; probe the representation separately.
Learns well from full examples early, then slows or stops improving with them Rising task-specific expertise making the guidance redundant; fatigue or loss of interest; a plateau on the target relation Give a completion problem or an unsupported changed problem; if independent success holds, reduce support provisionally. If it fails, the guidance may still be needed.
Lower effort with segmented or reduced material but no change in score Reduced extraneous processing without a learning gain; easier-looking material lowering investment; outcome measure insensitive to the change Hold the outcome measure fixed and add a delayed or near-transfer probe; report effort and performance separately rather than as one efficiency figure.

These rows are proposals, untested with learners. The probes can shift confidence among explanations; a single contrast does not identify a cause, and each probe is itself a learning exposure that can change the state.

Evidence and qualifications

  • Reducing extraneous load in multimedia design improved learning, more for complex material [+M]: an overview of 29 systematic reviews (1,189 studies, 78,177 participants) reports load-reducing design principles improving learning (g = 0.38), with g = 0.70 for complex, high-element-interactivity materials against g = 0.20 for simple ones. Learner prior knowledge did not significantly moderate effects, so it gives weak support to the expertise half of this model. The outcome is pooled learning (general-achievement), horizon not separated; it tests reducing load, not inducing overload. Not settled: the abstract available could not confirm its entries.
  • Integrating mutually referring text and diagrams beats separating them [+M]: a random-effects meta-analysis of 58 comparisons (n=2426) reports g = 0.63 for spatially integrated over separated presentation, holding broadly across moderators per the abstract. Learners' expertise and the horizon are not reported on the claim page. Not yet checked against its sources; it does not predict a benefit when each source is intelligible alone.
  • Isolating interacting elements first helped novices on complex tasks [+M]: one experiment reports that, for high-element-interactivity tasks, novices performed better when first given isolated parts before the integrated whole. Sample size, outcome measure and horizon are not reported on the page; not yet checked against its sources.
  • Guidance that helps novices can lose its advantage or reverse for experienced learners [~M]: three experiments with electrical trainees reading circuit diagrams found integrated text best for less experienced trainees and diagram-only best for the most experienced; a narrative review reports the same pattern across five load paradigms. Abstract only, no effect sizes or horizon recorded; not settled against its sources. It does not say at what level of expertise support should change.
  • Segmentation's efficiency advantage disappeared at higher prior knowledge [~M]: 75 secondary students, randomized to segmented or continuous animated probability examples. Segmented examples were more efficient (performance relative to mental effort) on near and far transfer at lower prior knowledge; the difference disappeared, not reversed, at higher prior knowledge, and there was no interaction on raw transfer performance. Not yet checked against its sources.
  • Lower-prior-knowledge learners did better with full than completion examples [~M]: a factorial experiment with college students on simultaneous equations; lower-prior-knowledge learners scored higher after full-worked examples (t(1,26) = 1.98, p = .05), while higher-prior-knowledge learners' advantage for completion examples was not significant (p = .22) and equivalence was not tested. Judge-checked against the full text; immediate posttest, no effect size printed.
  • Example-first sequences outperformed problem-first ones for novices [+M]: 96 Dutch secondary novices in circuit troubleshooting, randomized to examples only, example–problem pairs, problem–example pairs or problems only. The two example-first conditions scored higher on an immediate test (partial η² = .23) with lower rated training effort (partial η² = .20); problem–example pairs did not beat problems only, so starting with the example mattered. One study, immediate test only, narrow instruction-aligned items; equivalence between the two example-first conditions was not tested.
  • Contingent scaffolding improved learning more than fixed or absent support [+M]: fourth and fifth graders tutored one-to-one on long division (8 per condition) solved more problems at immediate and one-month tests under fully contingent support than under fixed high, moderate or no support; a meta-analysis of dynamic assessment ranks scaffolding above coaching but below explicit strategy training. Most of the evidence is one-to-one tutoring, so it bears on the response-dependent fading in this model, not on whole-class pacing.
  • The expertise reversal is more often partial than full [~M]: Kalyuga and Renkl's introduction to a special issue reports the reversal as replicated across many materials, usually as a partial reversal (a significant interaction, the novice advantage shrinking) rather than a full crossover. With the segmentation result above, it supports expecting support to stop helping before it starts to hurt; it gives no threshold.
  • Worked-example guidance becomes less effective as expertise grows [~M]: two syntheses (Kalyuga et al. 2003; Renkl & Atkinson 2003) argue that full examples should give way to problem solving as competence develops; neither prints an effect size, and the claim's own entries are abstract-level.
  • Chunking reduces working-memory load [+M]: list-recall experiments show that recoding items into units the learner already knows frees memory for other items held at the same time, and reviews place the limit for novel material near four chunks. No study on that page measures learning or problem solving after material was chunked for learners, so it supports the working-memory premise, not a design effect.
  • Reducing extraneous load can also remove affordances for germane processes [~W]: de Jong's 2010 critique of the theory, citing Paas et al. on hypertext, argues that turning a nonlinear environment into a linear one can cut the comparison and elaboration it prompted along with the load. An argument, no data; it is why the situation table asks whether the "extra" processing is the target before removing it.
  • Structured training to use generative AI as a partner raised scores without lowering rated load [+M]: Huang et al. (2026), a randomised trial with sophomore STEM students: training in using generative AI as a collaborative partner beat problem-solving training on an electromagnetism posttest (d = 0.81), while total rated load did not differ and intrinsic load was marginally higher [~M]. One study; it is the only test in the wiki behind the AI row, and it fits keeping the task's own demands with the learner rather than handing them to the tool.
  • Principles before a case comparison changed load ratings, not knowledge [~M]: Kienzler et al. (2023), student teachers learning operant conditioning; giving the principles before the comparison raised rated germane load (ηp² = .61) and lowered rated extraneous load (ηp² = .23), yet timing made no difference to the knowledge test [~M] (ηp² = .003). The reading of load ratings is held on Cognitive Load Management.
  • Duplicated words, and words beside pictures, are modelled on Redundancy and Modality: the rows above use voice with pictures beat printed text with pictures [+M] (36 adults), and two nulls for duplicate text, middle-school animation [~M] and a print probability lesson [~M]; a non-significant difference is not equivalence.

Keep learners, task, comparator, outcome and horizon with each result when transporting it. Efficiency, effort and performance are different outcomes; a non-significant difference is not equivalence; and the pooled values do not forecast an individual learner's response or rank these load-reducing techniques against each other.

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/cognitive-load.json, scripts/synthesize_claim.py). The count is of distinct studies across the claim and the narrower claims that bear on it.

Further evidence, not yet read against this model

Re-read 2026-10-19: thirteen of the sixteen claims listed here were moved into the model above (the situation table, the observation section and the evidence list). The three below were read and do not change the model.

Objective and learner-valued goal

Ask what the learner wants to accomplish and why, separately from the designer's objective. Agreement, divergence and uncertainty are observations; do not infer motivation from compliance or from a low effort rating. The designer chooses the representative task, the permitted aids, the criterion and the horizon; load reduction is a means to that objective, not the objective.

For example, an apprentice electrician may value reading a real wiring diagram on site without a reference card, while the designer's objective is passing a timed schematic-interpretation test. Integrated labels may raise the test score early yet leave the learner's valued performance (reading an unlabelled diagram) untested. Record both aims, and plan when the integrated support is removed and how unaided reading is checked.

What would revise this model?

The load-reduction expectation should weaken if comparable novices on high-element-interactivity tasks fail to benefit under defensible comparisons with aligned outcomes, or if benefits appear equally for low-element-interactivity material. The expertise qualification should weaken if, as in the multimedia overview above, prior knowledge repeatedly fails to moderate effects when it is measured. If effort falls without any gain in performance or transfer, the model's claim about learning (as opposed to comfort) is not supported for that design. If reducing load repeatedly costs learning where the removed processing was the comparison or integration the goal needed, the model should name which kinds of processing count as avoidable before it recommends removing them. Do not protect the model by relabelling every failure as "too much load" or "too much expertise" after the fact.

A learning-phase efficiency gain, immediate performance, near transfer, delayed retention and unaided use in the learner's setting are separate claims. The present evidence does not establish a threshold of expertise at which to remove support, a universal ordering of techniques, or that reduced load causes the gains it accompanies.

Claims

Examples

  • A novice algebra lesson uses a single integrated visual instead of separate text and diagram panels that learners must constantly coordinate.
  • A software onboarding flow replaces open-ended exploration with worked examples and progressively harder tasks until users can perform the workflow independently.
  • Example-problem sequences in math or programming.
  • Stepwise visual explanations paired with constrained early practice in technical training.
  • Instructional materials redesigned to remove split-attention and unnecessary visual clutter.
  • Manage intrinsic load with a part-whole approach, task variability, and self-explanation to raise germane load

Key Sources

  • Sweller, J., van Merriënboer, J. J. G., & Paas, F. (1998). Cognitive architecture and instructional design. Educational Psychology Review, 10(3), 251-296. https://doi.org/10.1023/A:1022193728205
  • Gupta, U., & Zheng, R. Z. (2020). Cognitive Load in Solving Mathematics Problems: Validating the Role of Motivation and the Interaction Among Prior Knowledge, Worked Examples, and Task Difficulty. European Journal of STEM Education, 5(1), 05. https://doi.org/10.20897/ejsteme/9252
  • Sweller, J., van Merrienboer, J. J. G., & Paas, F. (1998). Cognitive architecture and instructional design. Educational Psychology Review, 10(3), 251-296. https://doi.org/10.1023/A:1022193728205
  • Sweller, J. (2011). Cognitive load theory. Psychology of Learning and Motivation, 55, 37-76. https://doi.org/10.1016/B978-0-12-387691-1.00002-8
  • van Gog, T., Kester, L., & Paas, F. (2011). Effects of worked examples, example-problem, and problem-example pairs on novices' learning. Contemporary Educational Psychology, 36(3), 212-218. https://doi.org/10.1016/j.cedpsych.2010.10.004
Page metadata
Field Value
type principle
id cognitive-load-theory
aliases lower-extraneous-load-to-increase-germane-effort-in-stem
title Cognitive Load Theory
description For a task-specific novice on high-element-interactivity material, reducing avoidable processing is expected to improve performance or efficiency, a relationship that weakens or reverses as task-specific expertise grows.
status review
moved_from pattern/cognitive-load-reduction-clt-scaffolding-approach
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 4 — sweller-1998, gupta-2020, sweller-2011, van-gog-2011

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