Children become better at applying past experience to new situations as they grow older.
But new research suggests that this improvement may not come from one single developing ability.
A study published in Scientific Reports on 22 September 2026 tested 84 children aged 4, 6 and 8 years across four different forms of generalization: statistical learning, associative inference, transitive inference and categorisation.
Accuracy improved with age in every task. However, children who performed well on one task did not consistently perform well on the others.
That weak cross-task relationship raises an important question: rather than reflecting one unified “generalization ability”, these tasks may depend on partly different cognitive processes.
Computational modelling added another layer to the finding. Younger children often appeared to solve one inference task using simpler value-based learning, whereas 8-year-olds relied more heavily on explicit memory for directly learned relationships.
Generalization allows children to use old knowledge in new situations
Generalization is a broad term for the ability to apply what has been learned previously to something new.
A child who learns that two objects tend to appear together may later use that relationship to make a new inference. A child who recognises similarities among different animals may use those shared features to identify a new category member.
This ability is fundamental to learning because daily life rarely presents exactly the same situation twice.
Children constantly need to transfer knowledge from previous experiences to new people, objects, environments and problems.
The difficulty for researchers is that “generalization” is measured in many different ways.
Two tasks may both be described as testing generalization while relying on very different kinds of memory, attention or reasoning.
The researchers tested four forms of learning
Tydings M. McClary and colleagues tested 84 children divided among three age groups: 4-year-olds, 6-year-olds and 8-year-olds.
Each child completed several tasks designed to measure different forms of generalization.
The first was statistical learning, which involves detecting regularities or patterns across repeated experiences.
The second was associative inference. In this type of task, a child learns overlapping relationships and is later asked to infer a relationship that was never shown directly.
The third was transitive inference, where children learn ordered relationships and then infer how items relate when they were not directly compared.
The fourth was categorisation, which tests whether children can identify shared structure across different examples and apply it to new items.
Performance improved with age across all four tasks
The broad developmental pattern was clear.
Older children generally performed more accurately than younger children across the tasks.
The largest and most consistent differences appeared between ages 4 and 6.
On the tasks with sufficient data from 4-year-olds, the youngest group performed less accurately than both the 6- and 8-year-olds.
In several tasks, 8-year-olds also outperformed 6-year-olds.
This supports the broader idea that the capacity to use prior learning flexibly continues developing across early and middle childhood.
Doing well on one task did not reliably predict doing well on another
The more surprising result emerged when the researchers compared performance across tasks.
If all four paradigms measured the same underlying generalization ability, children who performed strongly on one task should also tend to perform strongly on the others.
The evidence for that pattern was weak.
Cross-task correlations were generally low or inconclusive.
This does not prove that the tasks are completely unrelated.
But it does suggest that performance may depend on different combinations of memory, pattern detection, reinforcement learning and explicit reasoning.
The finding therefore challenges the assumption that generalization should always be treated as one unitary cognitive skill.
Two inference tasks revealed especially different patterns
The researchers examined associative inference and transitive inference in more detail because both require children to derive a new relationship from previously learned information.
Despite that similarity, the two tasks produced different patterns.
In associative inference, children were generally better at remembering directly learned pairs than making new inferences.
In transitive inference, the reverse pattern appeared: inference performance could exceed memory for the directly learned pairs.
This difference suggests that children may reach correct answers in the two tasks through different mental routes.
Older children increasingly relied on explicit pair memory
Across both inference tasks, remembering the directly learned relationships became more useful with age.
For older children, successful inference was increasingly tied to accurate memory for the original pairs.
This suggests that as memory systems mature, children may become better able to deliberately retrieve specific learned relationships and combine them to generate a new conclusion.
By age 8, this pair-memory strategy appeared to play a stronger role than it did in younger children.
Younger children sometimes inferred correctly without remembering the original pairs
One of the most interesting findings came from the transitive-inference task.
Some 4- and 6-year-olds could make correct transitive inferences even when their memory for the directly learned pair relationships was relatively weak.
At first glance, this seems contradictory.
How can a child infer a new relationship if they cannot accurately report the information from which that inference should be built?
The answer may be that they were not solving the problem through explicit pair-by-pair memory at all.
Instead, they may have learned a simpler value structure across the items.
Computational modelling suggested a shift in strategy
The researchers used computational models to test different explanations for the transitive-inference results.
The modelling suggested that younger children were more likely to use a value-based reinforcement-learning strategy.
Rather than explicitly remembering every relationship, they may have gradually learned that some items were associated with higher or lower value based on previous feedback.
This could allow them to choose correctly on a new comparison even if they could not consciously retrieve all of the original pair relationships.
By age 8, the children were more likely to use a strategy based on explicit memory for the learned pairs.
The same correct answer can therefore emerge from different underlying processes at different ages.
A correct answer does not always reveal how the child solved the problem
This is one of the most important lessons from the study.
Behavioural accuracy alone can make two children look similar even when their underlying strategies are different.
A younger child and an older child may both select the correct answer, but one may rely on accumulated value signals while the other reconstructs the answer from explicit memories.
This matters for developmental research because improvement cannot always be understood simply by counting correct responses.
Researchers may need to ask not only whether children succeed, but how they succeed.
The findings could change how learning abilities are measured
Many psychological concepts are measured using one task and then treated as if that task captures the broader ability.
The new findings show why that can be risky.
If statistical learning, categorisation and different forms of inference rely on partly different mechanisms, one task may provide only a narrow view of a child’s broader learning abilities.
The authors therefore argue for multi-indicator approaches that combine several tasks rather than relying on a single paradigm.
Computational modelling may also help distinguish between strategies that produce similar outward performance.
The study does not show that one strategy is better
The age-related shift should not be interpreted as younger children’s strategy being simply inferior.
A reinforcement-learning approach can be efficient when direct memory is still developing.
It may allow children to extract useful structure from repeated experience without needing to consciously remember every relationship.
As memory systems mature, explicit retrieval may become more reliable and therefore more useful for complex reasoning.
The developmental change may therefore reflect an expanding set of available strategies rather than a simple replacement of a bad strategy with a good one.
There are important limitations
The study involved 84 children, which is modest for detecting small relationships between several cognitive tasks.
This means weak cross-task correlations should not be interpreted as definitive proof that the abilities are independent.
The authors themselves describe the evidence on cross-task concordance as inconclusive.
The categorisation task was also demanding enough that insufficient data were available from many 4-year-olds.
That limits direct comparison across all four tasks at the youngest age.
The computational modelling focused particularly on transitive inference, so the same strategy shift cannot automatically be assumed for every type of generalization.
The results do not translate directly into classroom instructions
The study was designed to understand cognitive development rather than to test teaching methods.
It therefore cannot tell teachers that one specific strategy should be used with 4-year-olds and another with 8-year-olds.
However, the findings do support a broader educational principle.
Children who arrive at the same answer may be using different mental processes, and younger learners may succeed without being able to explicitly explain or retrieve all of the information behind their decision.
Assessment that looks only at final accuracy may therefore miss meaningful differences in how learning is taking place.
The broader question is whether generalization is one ability at all
The study leaves its biggest question open.
Does generalization represent one cognitive ability that appears differently across tasks, or is it an umbrella term covering several distinct abilities?
The current results cannot settle that debate.
But they show that the answer cannot be assumed simply because several experiments share the same label.
Children clearly become more successful at applying prior experience to new problems between ages 4 and 8.
What changes underneath that improvement may be more complex: memory becomes more useful, strategies shift and different tasks may continue to depend on different mechanisms.
Source Information
Study Title: Generalization from early to middle childhood
Authors: Tydings M. McClary, Simon Ciranka, Elisa S. Buchberger, Bernhard Spitzer, Ulman Lindenberger, Chi T. Ngo and Markus Werkle-Bergner
Journal: Scientific Reports
Published: 22 September 2026
Sample: 84 children aged 4, 6 and 8 years.
Method: Children completed four generalization paradigms measuring statistical learning, associative inference, transitive inference and categorisation. The researchers compared age-related performance, examined correlations across tasks and used computational modelling to test alternative strategies underlying transitive-inference performance.
Main finding: Accuracy improved with age across all four tasks, but cross-task correlations were weak or inconclusive. Older children’s inference increasingly depended on memory for directly learned pairs, while modelling suggested younger children could solve transitive-inference problems using value-based reinforcement learning and 8-year-olds relied more on explicit pair-memory strategies.
DOI: 10.1038/s41598-026-69464-9
