HSW-0249 · How Studying Works
Two practice tasks can look as though they share the same mental machinery and still fail to transfer equally in both directions.
A learner may train a demanding task that requires continuously replacing outdated information while preserving what still matters. Performance on a simpler span task can then improve. Reverse the training direction—practise holding information without the same updating demand—and the improvement may not travel back upward as reliably.
That is working-memory transfer asymmetry: transfer depends not only on how similar two tasks are, but potentially on the direction in which their cognitive demands are related.
The point is not that “harder is always better.” The point is more precise. A 2026 systematic review and meta-analysis found reliable transfer from adaptive updating training to untrained span tasks, but not reliable transfer from span training to updating tasks. The pattern was small, heterogeneous and bounded by important methodological limits, yet it challenges the simple assumption that shared task features alone determine transfer.
This article owns the narrow study question of directional asymmetry in transfer between working-memory updating and span training. The broader canonical owner remains How Transfer of Learning Works. Brain Training for Exam Preparation retains the broad question of what cognitive training does and does not transfer to academic performance. Working-Memory Strategy Emergence retains the question of how better methods can raise working-memory-task performance without proving underlying capacity grew.
The 50-Second Read
- Transfer is directional. A skill that helps Task B after Task A does not guarantee that Task B will help Task A.
- Updating and span overlap but are not identical. Updating requires maintaining information while replacing what is no longer relevant; span tasks emphasize temporary storage and ordered access.
- A 2026 meta-analysis included 55 independent studies, 208 within-working-memory effect sizes and 3,492 participants.
- Updating → span transfer was reliable but small (Hedges’ g = 0.176), while span → updating transfer was not reliable (g = 0.048).
- The direct difference between directions was only marginal in the primary analysis. That matters: the pattern is evidence for asymmetry, not permission to exaggerate it.
- The asymmetry was clearer in some conditions, including non-adult samples and lower-dose studies, though moderator analyses had limitations.
- The educational implication is not “do n-back.” It is to design practice by asking what resources the future task includes, then verify transfer on the future task itself.
1. Similarity Is Not Direction
Traditional transfer explanations often begin with overlap.
If two tasks use similar operations, representations or knowledge, practice on one should help the other more than practice on a dissimilar task.
That is sensible, but it leaves a question unanswered:
If A and B overlap, must A → B transfer equal B → A transfer?
No. Overlap is symmetric as a description. Cognitive demand need not be.
2. Span and Updating Do Different Jobs
A span task asks a learner to preserve a set of items and reproduce them, sometimes while performing an interleaved operation.
An updating task asks the learner to maintain a changing state. Old information must be replaced, the currently relevant items must remain accessible, and interference must be controlled as the task evolves.
They share temporary memory demands, but updating can recruit additional control operations.
This creates two competing predictions. An ability-based account might expect span training to strengthen a broad maintenance component that updating also needs. A resource-based account predicts the opposite direction: training the more resource-demanding updating task may exercise a superset of processes sufficient for the lower-demand span task.
3. What the 2026 Meta-Analysis Actually Found
Fu and colleagues conducted a systematic review and three-level meta-analysis covering studies published through 30 September 2024. Their primary analysis included 55 independent studies, 208 effect sizes and 3,492 healthy participants who completed adaptive working-memory training.
The result was directional:
- updating training → untrained span tasks: g = 0.176, 95% CI [0.104, 0.249], p < .001;
- span training → untrained updating tasks: g = 0.048, 95% CI [-0.078, 0.173], p = .453.
The direct contrast between the two directions in the primary moderator analysis was marginal rather than conventionally significant, F(1,202) = 3.259, p = .073. That statistical detail is important. The reliable effect in one direction and absent reliable effect in the other are not, by themselves, proof that the two effects differ. The authors therefore add moderator, sensitivity and extension analyses rather than resting the interpretation on a slogan.
Read the open-access source: Fu et al., 2026, The asymmetry of working memory training transfer.
4. The Effect Is Small
Hedges’ g = 0.176 is not a transformation of general intelligence.
It is a small average transfer effect across heterogeneous studies.
This distinction protects learners from a common mistake: taking a statistically reliable pattern and inflating it into a life-changing study hack.
The correct reading is:
Across these training studies, updating practice produced a small reliable average transfer to span tasks; span practice did not show reliable average transfer to updating tasks.
That is already interesting. It does not need to be exaggerated.
5. Why a Superset Can Transfer Downward
Imagine two procedures.
Task A requires: hold four items.
Task B requires: hold four items, continuously replace selected items, suppress obsolete states and preserve the correct current configuration.
If the second task genuinely recruits the first task’s resources plus additional control, training B may improve performance components useful for A.
Training A does not necessarily train the extra operations demanded by B.
This is the resource-based interpretation. It is a theoretical account supported by the observed pattern, not a direct measurement of a universal “resource hierarchy.”
6. More Demanding Is Not Automatically Better
A dangerous educational translation would be:
Make every exercise harder than the examination and transfer will take care of itself.
That does not follow.
Extra difficulty helps only when the additional demand trains processes that the future task genuinely needs. Difficulty caused by confusing instructions, irrelevant switching, sensory noise or overloaded notation can consume resources without strengthening the target capability.
The right question is not “Is this harder?” It is:
What extra operation does this practice require, and will that operation still matter later?
7. Moderator Findings: Where the Asymmetry Looked Stronger
The review examined training dosage, developmental stage and stimulus domain.
In non-adults, updating-to-span transfer was g = 0.362 while span-to-updating transfer was not reliable; the difference between directions was statistically supported. In adults, updating-to-span remained reliable, but the directional difference was not statistically supported.
The authors also reported clearer asymmetry in lower-dose studies and studies using verbal materials. Yet several moderator cells contained relatively few studies, and the authors caution that imbalance and measurement sensitivity can affect these comparisons.
Moderator patterns are boundary clues, not new universal laws.
8. Transfer Beyond Working Memory
The researchers also examined non-working-memory outcomes such as processing speed, inhibition, shifting, mathematics, intelligence-related measures and academic achievement.
Updating training showed a small reliable average effect, g = 0.120, whereas span training did not. Again, the difference between directions was marginal rather than decisive in the primary extension analysis.
This is not strong evidence that updating training broadly raises school achievement. The outcomes were heterogeneous, and far-transfer claims require especially careful interpretation.
9. Mathematics Example: Train the State Change, Not Only the Snapshot
An illustrative algebra learner can solve a single substitution when all values are visible but loses track during a multi-step derivation.
Simple practice may ask the learner to hold one intermediate value.
A richer task may require the learner to update the current equation after each valid transformation while discarding the obsolete state.
If the examination requires evolving state control, practice that includes state updating is closer to the real demand than practice that merely increases the number of static items.
This is an educational analogy, not direct evidence that algebra practice follows the exact meta-analytic working-memory mechanism.
10. English Example: Revision Is an Updating Task
Editing a paragraph is not just holding the original wording.
The writer must preserve the argument, replace weak wording, update pronoun references, remember what has already changed and prevent an old version from re-entering.
Practising isolated sentence recall may strengthen temporary retention. It does not automatically train the dynamic control required to revise a live text.
11. Science Example: Updating a Model After New Evidence
A science learner may hold a hypothesis perfectly and still fail to revise it when new evidence arrives.
Good reasoning requires a state transition:
old model → new evidence → preserve valid parts → replace invalid parts → updated model.
That is structurally closer to updating than to passive retention.
12. Working-Memory Transfer Asymmetry vs Brain Training
Brain Training for Exam Preparation owns the broad question of whether cognitive exercises transfer to school performance.
This article takes a narrower job: when two working-memory paradigms share processes, does transfer differ by direction?
The answer from this 2026 synthesis is: apparently yes on average, but modestly and with important limits.
13. Transfer Asymmetry vs Context-Carried Strategy Transfer
Context-Carried Strategy Transfer asks whether a control strategy travels because environmental cues overlap.
Transfer asymmetry asks a different question: even when two task families overlap, is one training direction more capable of supplying what the other task needs?
14. The Resource-Superset Audit
Before replacing ordinary practice with a more demanding variant, write down the resource structure.
| Future task demand | Present in current practice? | Present in harder practice? |
|---|---|---|
| Maintain current state | Yes | Yes |
| Replace obsolete state | No | Yes |
| Suppress interference | Sometimes | Yes |
| Select next operation | No | Yes |
| Irrelevant visual clutter | No | Yes |
The first four rows may build a useful superset. The last row may merely add noise.
15. The Direction Test
- Choose Task A and Task B.
- List the processes each requires.
- Mark which processes are shared.
- Mark which processes exist only in A or only in B.
- Predict A → B and B → A separately.
- Train one direction.
- Test the untrained task after a delay.
- Do not infer symmetry unless both directions were actually tested.
16. The Delayed Independent Check
Training performance is not transfer.
After practice, remove the training interface and wait. Then test a new task that requires the proposed shared process without copying the trained surface.
Ask three questions:
- Did performance improve on the trained task?
- Did performance improve on the untrained target?
- Did the advantage survive after the training cues disappeared?
Only the second and third questions speak directly to transfer.
17. Parent and Tutor Guide: Do Not Buy “Brain Improvement” From a Near-Transfer Result
If a programme claims to improve working memory, ask what the child became better at after training.
- the exact game?
- a similar working-memory task?
- a different working-memory paradigm?
- mathematics or reading?
- real examination performance?
These are different claims. A reliable small transfer between updating and span tasks does not justify treating a cognitive-training product as a substitute for subject knowledge, retrieval practice, worked examples, feedback or authentic examination training.
18. For Teachers: Build the Future Operation Into Practice
Rather than asking whether an exercise is generally “cognitively demanding,” ask which operation the final performance will require.
- If the final task requires only recall, practise recall.
- If it requires recall plus updating, include updating.
- If it requires method selection, include method competition.
- If it requires revision after feedback, include live revision.
- If it requires switching between representations, practise the switches.
The more honestly practice represents the target operation, the less you depend on hoped-for transfer.
19. What Not to Conclude
- Do not conclude that difficult practice always transfers downward.
- Do not conclude that span practice is useless; it may improve span performance and other outcomes under specific conditions.
- Do not interpret a small average effect as a guaranteed individual benefit.
- Do not infer academic transfer from near transfer inside working memory.
- Do not treat non-significance in one direction as proof of exactly zero effect.
- Do not ignore the marginal primary directional contrast.
- Do not assume the moderators establish causal developmental mechanisms.
20. Evidence Boundary
The meta-analysis included adaptive training studies of healthy participants, drew from English- and Chinese-language databases and covered literature through September 2024. The authors report substantial between-study heterogeneity, unequal numbers of studies across transfer directions and limited data in some moderator cells.
Its strongest educational contribution is therefore conceptual rather than prescriptive: transfer can be asymmetric, and a task’s resource relationship to another task may matter in addition to surface similarity.
21. The Training Rule
When preparing for a demanding performance, do not merely practise something that looks similar.
Map the operations.
Identify the future bottleneck.
Practise a task that actually recruits it.
Then test whether the capability travels.
22. Return: Transfer Has an Arrow
Transfer is often drawn as a bridge between two tasks.
A better picture includes an arrow.
What A gives B may not equal what B gives A.
Study design improves when we stop asking only whether two tasks are similar and start asking what one task trains that the other will later require.
Continue through How Transfer of Learning Works, Brain Training for Exam Preparation, Working-Memory Strategy Emergence, the How Studying Works Numbered Series Reading Index and the How X Works Hub.