HSW-0301 · How Studying Works
“Use retrieval practice” is good advice often enough that it can become dangerous when turned into a law.
A student learns historical causes, vocabulary, scientific definitions and mathematical procedures by trying to retrieve them before checking. The method works well. The student then assumes the same rule must apply to every kind of information: unfamiliar diagrams, abstract shapes, arbitrary visual textures, exact colour-pattern pairings, spatial configurations and every other thing that can possibly be remembered.
That assumption is stronger than the evidence.
Retrieval practice has an important material boundary: a method that reliably improves many kinds of meaningful knowledge does not automatically produce the same advantage for every representation a learner might need to remember.
This article owns that boundary. It does not replace the general mechanism in How Retrieval Practice Works. It does not replace Retrieval-Practice Generalisability, which asks whether a strong average effect can be assumed for every learner. Here the question is different: does the kind of material itself change whether retrieval beats restudy?
Quick Answer
Yes. The evidence base for retrieval practice is large, but it is not representation-free. A 2025 six-experiment paper in the Journal of Experimental Psychology: Learning, Memory, and Cognition found no testing advantage for novel, meaningless visual objects. Across several attempts to make retrieval practice more favourable—including feedback, delayed testing, changed choice formats and drawing—the tested objects were remembered about as well as, or sometimes worse than, restudied objects. See McCarter, Huber and Cowell (2025).
A separate 2026 study in health-professions education also found that retrieval-enhanced learning varied across topics: testing improved one-week retention in most comparisons, but the size of the benefit was not uniform and one content set did not show the expected advantage. See Scott and colleagues (2026).
Neither result overturns retrieval practice. They sharpen the rule: test the capability you need, but do not assume the testing effect has the same magnitude for every material, representation, task or outcome.
1. Retrieval Practice Is a Comparison, Not a Magic Action
The phrase “retrieval practice works” compresses an experimental comparison. Researchers usually compare some form of retrieval attempt with another learning condition such as restudy, rereading or no additional practice.
That means two questions matter:
- Does retrieval produce learning?
- Does retrieval produce more useful later performance than the alternative use of the same study opportunity?
A learner can improve during retrieval practice while restudy improves even more. The existence of learning is not the same as a testing-effect advantage.
2. Meaning Gives Retrieval More Than One Route Home
Many classic retrieval-practice materials are words, facts, concepts, passages or paired associates with semantic structure. A learner retrieving “mitochondrion” can use meaning, category, sentence context, related ideas, prior knowledge and language.
A novel abstract object made from an unfamiliar outline and unfamiliar fill pattern may offer far fewer pre-existing routes. The learner can still encode and retrieve it, but the operation is not cognitively identical to recovering a meaningful concept from a network of knowledge.
The 2025 visual-testing paper explicitly raises this possibility. Its authors do not prove that semantic content is the missing ingredient, but they show that a strong testing advantage did not emerge when meaning and familiar labels were deliberately stripped away.
3. Six Experiments Matter More Than One Null Result
A single failed experiment could reflect a weak manipulation. McCarter and colleagues therefore changed the procedure repeatedly. They varied delay, feedback, practice choices and even used drawing as retrieval practice.
The important result is not “visual memory cannot benefit from retrieval”. The study does not justify that universal statement. The defensible result is narrower: under the tested conditions, purely visual practice with novel meaningless objects did not show the familiar retrieval-over-restudy advantage.
That is exactly the kind of boundary an evidence-led study system should preserve.
4. Do Not Turn a Boundary Finding Into the Opposite Myth
The wrong reaction is to say, “Retrieval practice is only for words.” Diagrams, images and visual-spatial material can contain meaning, labels, structure, relations and familiar components. Drawing from memory can also perform valuable diagnostic and generative jobs.
The boundary concerns the match among material, representation, retrieval operation and outcome—not a crude verbal-versus-visual division.
5. The Unit of Retrieval Matters
Suppose a Biology learner studies a labelled heart diagram. There are several possible retrieval jobs:
- name a structure from its location;
- place a label on a blank diagram;
- draw the rough structure from memory;
- explain blood flow through the chambers;
- identify what changes when a valve fails;
- distinguish one structure from a visually similar neighbour.
“Retrieve the diagram” is therefore underspecified. Different tests strengthen and diagnose different access routes.
6. Meaningful Visuals Are Not Meaningless Objects
A geometry diagram, chemical structure, graph, map or anatomical drawing often carries relational meaning. The learner can encode position, function, label, quantity and causal relation.
It would be a mistake to apply a null result from novel meaningless objects directly to all educational images. The experiment is valuable precisely because it tells us not to overgeneralise in either direction.
7. Topic Can Matter Even Inside Meaningful Education
The 2026 health-education study compared retrieval with extended study across several content sets and outcome measures. Retrieval advantages appeared in most comparisons, but not all, and exploratory analyses did not fully explain why the magnitude varied by topic.
This should increase precision, not pessimism. A robust average method can still have topic-dependent effect sizes.
Use retrieval because it has strong evidence—not because strong evidence means every topic must respond identically.
8. The Reader Job Is Method Selection Under Uncertainty
When a learner meets a new kind of material, the decision is not “retrieval or never retrieval”. It is:
- What representation must survive?
- What future task will demand it?
- What retrieval operation reproduces that demand?
- What alternative practice would occupy the same time?
- What delayed evidence would tell us which method worked better?
That turns a study slogan into a controlled learning decision.
9. A Mathematics Example: Formula, Shape or Structural Relation?
Imagine a student learning transformations. If the future task is to recall a rule, verbal or symbolic retrieval may be ideal. If the future task is to recognise a rotation from an unfamiliar diagram, repeated classification and discrimination may be more important. If the future task is to construct the image, the learner needs production practice.
The correct study operation follows the required capability, not the prestige of one technique.
10. A Science Example: Remembering a Micrograph
A learner studying microscopy images may remember the verbal label while failing to recognise the structure when scale, orientation or contrast changes.
Blank-page recall of the label does not solve that. The learner needs repeated visual discrimination across varied authentic instances and then independent classification. Retrieval remains part of the system, but the target representation has changed.
11. An English Example: Quotation Versus Argument
A literature learner might need exact wording for a quotation and gist-level understanding for the argument surrounding it. Free recall of the theme cannot substitute for exact verbal memory when wording matters. Conversely, memorising a sentence perfectly does not prove the learner can explain its significance.
The material boundary intersects with the performance boundary.
12. Retrieval Format Can Rescue or Distort the Test
Recognition, cued recall, free recall, reconstruction and drawing are different operations. A learner who fails free drawing may still have useful recognition memory; a learner who recognises an image may still be unable to reconstruct it.
See Retrieval Format for the broader distinction. Here the practical rule is to avoid declaring a material “not learned” from one mismatched retrieval format.
13. Feedback Does Not Guarantee a Testing Advantage
Feedback is usually valuable after an unsuccessful retrieval attempt. Yet the 2025 visual study included feedback manipulations and still did not produce a reliable retrieval-over-restudy advantage.
This is another useful correction to slogan thinking: adding a generally helpful ingredient does not guarantee that the complete method wins in every representation space.
14. A Delayed Test Is Still Essential
Immediate practice performance is a poor final judge. A method can look hard now and pay later, or look fluent now and collapse later.
For unfamiliar materials, compare methods on a delayed test after enough time for short-lived practice states to fade.
15. Use a Changed-Representation Test
If the learner studies one diagram, do not test only the identical diagram. Rotate it. Remove labels. Change superficial features. Ask for the same relation in a new instance.
This distinguishes memory for the studied surface from a representation that can support transfer.
16. The Restudy Straw Man Problem
“Restudy” is often treated as passive staring. It need not be. A careful restudy pass can compare features, trace relations, align examples and notice distinctions missed earlier.
When retrieval is compared with a strong alternative, the difference may be smaller than when it is compared with weak rereading. That does not diminish retrieval practice; it improves experimental and study-method honesty.
17. Build a Material-Specific Study Pair
Instead of choosing one method permanently, pair study operations:
- Meaningful facts: retrieval + feedback + spacing.
- Procedures: retrieval of decisions + execution on fresh problems.
- Visual discrimination: varied examples + forced distinctions + explanation.
- Exact visual construction: observation + reconstruction + overlay comparison.
- Integrated diagrams: drawing + labelling + causal explanation.
The pair is allowed to change as the learner changes.
18. Do Not Generalise From One Learner Either
Material boundaries and learner heterogeneity can interact. A diagram may be meaningful to an expert and nearly meaningless to a novice. A notation system that compresses information for one learner may look like arbitrary marks to another.
The same nominal stimulus can therefore occupy different knowledge states.
19. AI Can Hide the Representation Problem
An AI tool can convert a visual object into a verbal explanation, label a diagram or describe a pattern. That may help build meaning—but it also changes the task.
If the learner ultimately needs unaided visual recognition, the verbal explanation cannot be the only practice. Ask who performed the target discrimination and whether the capability survives without the tool.
20. Diagnostic Questions
- Is the material meaningful or largely arbitrary to this learner?
- Does the learner have labels or schemas that organise it?
- Is the future task recall, recognition, discrimination, construction or explanation?
- Does retrieval practice reproduce the future cognitive operation?
- Would a strong restudy/comparison pass teach a missing relation better?
- What happens after delay?
- What happens when surface features change?
21. Parent and Tutor Guide
If a child says flashcards “do not work” for a topic, do not jump to a fixed learning style. Inspect the representation. Flashcards may be excellent for definitions and poor for learning to read an unfamiliar map, interpret a complex diagram or distinguish subtle visual patterns.
Change the practice operation before changing the learner story.
22. Evidence Boundary
The 2025 novel-object experiments are strong evidence against assuming a universal testing advantage under those conditions. They are not proof that retrieval is ineffective for all visual learning. The 2026 topic study shows effect-size variation across educational content but does not establish a simple rule for predicting which topic will benefit most.
Both sources therefore support a disciplined conclusion: material properties can matter, the boundary is not fully mapped, and learners should preserve retrieval practice while testing its fit rather than turning it into doctrine.
23. The Material-Boundary Protocol
- Name the representation that must survive.
- Choose a retrieval operation that actually samples it.
- Choose a strong comparison method, not a straw-man alternative.
- Keep practice time approximately comparable where practical.
- Test after delay.
- Change cues, examples or orientation when transfer matters.
- Measure the target capability, not merely completion.
- Keep the method only if the evidence justifies the extra effort.
24. Return: Robust Does Not Mean Universal
Retrieval practice deserves its place among the strongest general study methods. Its strength is not weakened by discovering boundaries. Scientific confidence becomes more useful when it has edges.
Retrieve meaningful knowledge aggressively. For unfamiliar representations, test what retrieval is actually doing. When the material changes, allow the method mix to change too. The goal is not loyalty to retrieval practice. The goal is durable capability.
Continue through Retrieval-Practice Generalisability, Retrieval Format, Drawing-to-Learn Integration, the How Studying Works Numbered Series Reading Index and the How X Works Hub.