HSW-0179 · How Studying Works
Retrieval practice is one of the strongest tools in the study toolbox.
But strong tools do not pay the same return everywhere.
Imagine two cue–answer pairs.
- bird → robin
- harbour → arithmetic
The first cue already points strongly toward a related answer. The second does not.
If you test both pairs instead of restudying them, will retrieval practice add the same amount of learning?
A 2026 experiment suggests the answer can be no.
Semantic relatedness is the degree to which the meaning of a cue already supports or predicts its target. When that relation is strong, retrieval practice can still help, but the extra benefit over restudy may be smaller because semantic knowledge is already providing a powerful route.
This article owns the narrow problem of semantic relation strength as a moderator of the incremental testing benefit. The broader mechanism and use of practice testing remain owned by How Retrieval Practice Works. The vocabulary-specific risks of learning related words together remain owned by Semantic Grouping Interference.
The 50-Second Read
- Retrieval practice still worked. In a 2026 study, test-with-feedback improved 24-hour cued recall for both weakly and strongly related word pairs.
- The incremental gain was not equal. The testing benefit was about 26% smaller for highly related pairs under the study’s modelling approach.
- Strong semantic scaffolds can support both study and test. That may reduce how much additional memory strength testing contributes over restudy.
- This is not a reason to stop testing related material. The effect remained positive.
- It is a reason to allocate retrieval intelligently. Spend more testing effort on weak links, arbitrary relations, exceptions, discriminations and material that cannot be reconstructed easily from meaning alone.
- Do not generalise from word pairs to every school subject. The experiment used controlled cue–target materials; educational transfer requires verification.
1. Retrieval Practice Has an Opportunity-Cost Problem
A learner has twenty minutes and sixty items.
One option is to test everything equally.
Another is to ask where retrieval practice buys the most improvement relative to easier alternatives.
That second question matters because some knowledge already has strong support from meaning, structure, prior knowledge or predictable relations. Other knowledge is arbitrary, weakly connected or vulnerable to confusion.
The value of retrieval practice is not only whether it works. It is how much extra learning it produces in this location of the knowledge network.
2. What the April 2026 Study Tested
Gupta, Pan and Rickard published “Semantic relatedness and the efficacy of retrieval practice” in npj Science of Learning on 4 April 2026.
The researchers recruited university students to learn English word pairs that were either weakly or strongly semantically related. During training, some pairs were restudied and others were tested with feedback. Final cued recall occurred 24 hours later.
The final analysed sample included 231 participants: 118 in the low-relatedness group and 113 in the high-relatedness group.
Testing improved recall in both relatedness conditions. However, using the authors’ process-model and distribution-matching approach, the incremental testing benefit was about 26% smaller for the highly related pairs. The authors reported that this pattern was not explained by simple ceiling constraints.
3. The Important Sentence Is “Testing Still Helped”
The wrong headline would be:
Do not use retrieval practice for related material.
That is not what the study found.
Testing improved later recall in both groups. The question was about relative efficacy: how much additional benefit testing produced compared with restudy under different semantic conditions.
For students, this distinction is crucial. A method can be beneficial and still have a lower marginal return in one part of the curriculum than another.
4. Why Strongly Related Material May Need Less Extra Help
When a cue and target are strongly related, pre-existing semantic knowledge already creates a scaffold.
The cue doctor makes medically related concepts more accessible than an arbitrary target. Study and testing can therefore draw on overlapping semantic support.
The 2026 paper develops a modelling account in which memory strengths formed during study and testing may become positively correlated for highly related material because both draw from shared semantic features.
That is one plausible explanation for the reduced incremental benefit. It is not the only imaginable mechanism, and it should not be turned into a universal classroom law.
5. Semantic Support Is a Kind of Built-In Cueing
Consider learning that evaporation involves liquid becoming gas.
The relation is meaningful and mechanistically connected. Many later questions can be reconstructed from the concept network.
Now consider an arbitrary historical date, an exception to a spelling pattern, a scientific constant or the name of an unfamiliar anatomical structure.
Meaning may provide less support. Those targets can depend more heavily on deliberately strengthened cue–target routes.
This does not mean “memorise arbitrary things, understand meaningful things.” It means the existing semantic scaffold is part of the learning environment and should affect how practice time is allocated.
6. The Difference Between Absolute Benefit and Incremental Benefit
Suppose restudy produces 70% recall and testing produces 85%. Testing helped by 15 percentage points.
Elsewhere, restudy produces 40% and testing produces 65%. Testing helped by 25 points.
Testing is useful in both places, but the second location offers a larger increment.
The numbers here are illustrative, not results from the 2026 study. The principle is the important part: method effectiveness should be evaluated against what would have happened otherwise.
7. Mathematics: Strong Structure Can Reconstruct What Weak Memory Forgets
Some Mathematics facts are strongly recoverable from structure.
If a student understands why the gradient of a perpendicular line relates to the negative reciprocal under the usual coordinate geometry conditions, the relation can be reconstructed from a broader conceptual network.
Other elements may be more arbitrary:
- notation conventions;
- special values;
- domain restrictions;
- which formula is supplied and which is not;
- rare exception cases.
A smart revision system does not test only what feels difficult. It asks which knowledge has the weakest alternative route home.
8. English: Semantic Neighbours Can Help and Compete
Vocabulary provides an obvious application, but it also introduces another problem.
Related words can support meaning while also increasing competition. Reluctant, hesitant, unwilling and resistant occupy nearby semantic territory but are not interchangeable.
That is why Semantic Grouping Interference remains a separate canonical problem.
For related vocabulary, retrieval practice should increasingly test discrimination: not only “What does this word mean?” but “Why is this word better than its neighbour here?”
9. Science: Test the Weak Link in the Causal Chain
A science explanation can contain strongly connected steps and one weak step.
For example:
temperature rises → particles move faster → collisions become more frequent → successful collision rate changes → reaction rate changes
A learner may retrieve the first and final statements because they are familiar while losing the condition that links collision energy to reaction success.
Retrieval practice should therefore target the least semantically inevitable step, not merely repeat the whole chain equally.
10. Semantic Relatedness vs Retrieval Practice Itself
How Retrieval Practice Works owns the broad question of why actively bringing knowledge to mind can improve later learning compared with passive review.
This article asks a narrower optimisation question:
Does the semantic relationship between cue and target change how much extra benefit testing provides?
The 2026 evidence says it can, at least in the tested word-pair paradigm.
11. Semantic Relatedness vs Cue Overload
Cue Overload concerns one cue pointing to too many targets.
Strong semantic relatedness can make one cue-target relation easy, while a dense semantic neighbourhood can create competition among several related targets. These are different properties.
Do not assume “more related” always means “easier.” Relatedness can support association and simultaneously demand finer discrimination.
12. Semantic Relatedness vs Learning Discrimination
Learning Discrimination owns the job of telling similar-looking problems or concepts apart.
Semantic relatedness can create the conditions under which discrimination becomes necessary. If meaning already brings several neighbours into reach, retrieval practice should sometimes test the choice among neighbours rather than only the availability of the general semantic field.
13. Build a Semantic-Support Map
Take a topic and classify what must be learned.
| Knowledge type | Semantic support | Likely practice need |
|---|---|---|
| Core mechanism | High if understood | Explain + retrieve across cues |
| Arbitrary label | Low | Repeated retrieval + spacing |
| Exception | Often low or misleading | Contrastive retrieval |
| Closely related alternatives | High but competitive | Discrimination practice |
| Multi-step procedure | Variable | Retrieve decisions, not only sequence |
This is a planning aid, not a fixed taxonomy. The same item can move categories as understanding changes.
14. Retrieval Should Target What Meaning Cannot Safely Supply
If a concept can be reconstructed reliably from a mechanism, do not stop retrieving it altogether.
Instead, spend disproportionate retrieval effort on:
- weak cue-target relations;
- arbitrary details;
- exceptions;
- confusable alternatives;
- boundary conditions;
- steps that are repeatedly skipped;
- knowledge that must be produced quickly under pressure.
This converts retrieval practice from a ritual into allocation.
15. Do Not Let Strong Relatedness Hide False Knowledge
A strong semantic cue can make an answer feel inevitable.
That creates a verification risk.
Students can produce a plausible related answer that is not the target. In Biology, respiration can cue oxygen even when anaerobic respiration is the relevant case. In Economics, inflation can cue prices without the learner distinguishing a general price-level change from one product becoming expensive.
High semantic relatedness can therefore support recall while demanding stricter precision.
16. The Center-to-Edge Retrieval Portfolio
- Center: retrieve the core concept and mechanism.
- Near ring: retrieve strongly related examples rapidly.
- Contrast ring: distinguish neighbouring concepts.
- Weak-link ring: retrieve arbitrary, low-association and exception material.
- Edge: solve unfamiliar tasks where the semantic relation is not given in the wording.
This retains retrieval practice while reallocating its intensity.
17. The School Route: Do Not Give Every Fact the Same Testing Budget
A quiz can contain twenty items and still sample learning badly.
If most questions ask for obvious semantic associates, performance may look strong while arbitrary details, exceptions and discriminations remain weak.
Teachers can build a better retrieval set by mixing:
- high-relatedness core knowledge;
- weakly cued details;
- near-neighbour choices;
- changed wording;
- application requiring method selection.
18. The Systems Route: Redundancy Changes Marginal Value
Systems become robust when important functions have multiple routes.
Semantic knowledge can act as one route. Episodic memory from study can act as another. A retrieval-practice episode may add another trace or access route.
When several routes already overlap strongly, another investment can still help but may add less. When a target has one fragile route, another route can be disproportionately valuable.
19. The Financial Route: Allocate Practice by Marginal Return
Study time is scarce capital.
The right question is not “Does retrieval practice work?” It is also:
Where will the next retrieval attempt buy the largest useful improvement?
This is marginal-return thinking. A well-known core fact may need maintenance. A weak prerequisite may need repeated retrieval. An exception may need contrast. A secure fact may need transfer rather than another identical flashcard.
20. The Learning Route: Use Feedback to Stop Plausibility From Masquerading as Recall
Related cues make plausible guesses easier.
That makes feedback especially important. After retrieval:
- compare the exact target;
- mark partial answers;
- identify whether meaning supplied a guess;
- retrieve the corrected answer again later.
“Close enough” can be a dangerous scoring rule when neighbours matter.
21. The Education Route: Teach Students Why Methods Have Different Marginal Value
Students often search for one best technique.
There is a better question:
What weakness is this technique supposed to repair?
Retrieval practice is highly useful, but its job can change:
- strengthen weak access;
- maintain durable knowledge;
- expose overconfidence;
- train discrimination;
- test transfer;
- verify independent production.
Technique choice should follow the learning problem.
22. The Training Route: Low-Association First Pass
- List the knowledge in a topic.
- Mark items that are easily reconstructed from meaning.
- Mark items that feel arbitrary or weakly cued.
- Mark groups that are strongly related but easily confused.
- Retrieve the weakly cued items first.
- Use contrastive retrieval for confusable groups.
- Return to core related knowledge through changed questions.
- Retest the whole set after delay.
This keeps the network coherent while concentrating effort where semantic support is weakest.
23. The Improvement Route: Compare Gains, Not Just Scores
If a learner scores 95% after testing, that may look excellent.
But if restudy would have produced 94%, the method added little in that specific case. If another weak set moves from 45% to 70%, the same amount of practice may have produced much more learning value.
In real study you rarely know the counterfactual exactly. Use small within-learner comparisons cautiously: similar material, similar delays, changed methods, repeated enough times to avoid overreacting to one result.
24. The World Route: Experts Know What Can Be Reconstructed and What Must Be Retrieved Exactly
Professionals do not memorize every detail equally.
They internalise core structures, retrieve safety-critical and high-frequency facts rapidly, verify details that should not be guessed, and know when a relation is strong enough to reconstruct versus when precision requires reference.
Study should develop the same judgment: what meaning supplies, what memory must supply, and what external reference should supply.
25. Parent and Tutor Guide: Ask What Is Hard to Rebuild
When deciding what a learner should retrieve repeatedly, ask:
- Can the learner reconstruct this from a secure principle?
- Is it arbitrary?
- Is it easily confused with a neighbour?
- Does it need rapid access?
- Is it a prerequisite for later work?
- Does the student repeatedly guess a plausible but wrong related answer?
The answers help allocate retrieval rather than simply increase it.
26. What Not to Do
- Do not stop retrieval practice for strongly related material; the 2026 study still found a benefit.
- Do not turn “26% smaller” into a universal percentage for classrooms or subjects.
- Do not assume semantic relatedness always makes learning easier; competition can rise too.
- Do not spend equal retrieval time on every item by default.
- Do not confuse plausible guessing with precise recall.
- Do not infer that arbitrary facts are more educationally important merely because retrieval practice adds more.
27. Evidence Boundary
The 2026 study used English cue–target word pairs in a controlled laboratory design with university students. Semantic relatedness was operationalised using normative forward associative strength. The finding that testing’s incremental benefit was smaller for strongly related pairs is therefore evidence about that paradigm, not proof that every conceptually related school topic will show the same percentage reduction.
Its broader value is to demonstrate that retrieval-practice efficacy can depend on the structure of the material. Educational applications should preserve the proven broad value of retrieval while testing how allocation, cueing and discrimination affect real learners and real subject matter.
28. Return: Test What Meaning Cannot Reliably Carry for You
Retrieval practice remains powerful.
The more advanced question is where its next unit of effort is worth most.
Maintain the strong semantic core. Retrieve weak and arbitrary links more aggressively. Contrast close neighbours. Test exceptions. And use delayed performance to find out whether meaning is genuinely supporting memory or merely making the answer feel obvious.
Continue through How Retrieval Practice Works, Cue Overload, Learning Discrimination, the How Studying Works Numbered Series Reading Index and the How X Works Hub.