Mutualism Theory of Intelligence: How Cognitive Abilities May Compound

The familiar g factor is a robust statistical summary: performance on different cognitive tasks tends to correlate positively. The open question is what developmental machinery produces that pattern.

There are several possibilities. A shared biological mechanism — for example faster neural conduction or more efficient information processing — might benefit many tasks. A process-overlap account instead emphasises domain-general processes such as working memory and attention control that are recruited across otherwise different tasks. Mutualism adds a developmental possibility: abilities that begin partly distinct may become correlated because growth in one changes the conditions for growth in another.

A concrete example is vocabulary and non-verbal reasoning. A child with richer vocabulary may understand instructions, categories and relations more efficiently, giving them more productive reasoning practice. Better reasoning can in turn make it easier to infer new word meanings from context. Longitudinal evidence suggests that these two abilities can indeed predict one another's later growth.

The stronger proposal developed here is asymmetric developmental mutualism: some cognitive relationships may be reciprocal, some mainly one-way, many weak, and some dependent on the cognitive niche — the recurring tasks, tools, people and challenges through which a person's abilities are exercised.

The positive manifold is one of the most replicated findings in intelligence research. Someone who performs relatively well on matrix reasoning is, on average, also more likely to perform relatively well on vocabulary, working-memory and processing-speed measures. Factor analysis therefore extracts a broad general intelligence factor, g.

This article does not require denying that psychometric fact. Cognitive abilities are also substantially stable. A 2024 meta-analysis of 205 longitudinal studies involving 87,408 participants found that rank-order stability rises rapidly through childhood and is consistently high from late adolescence onwards. At age 20, over a five-year interval, average observed stability was about .76.

The deeper issue is explanatory. A covariance pattern does not uniquely tell us what generated it. Psychometric g may be useful for prediction even if the underlying causal system is distributed across many interacting biological and cognitive processes.

Useful distinction: psychometric g describes broad covariance. A theory of intelligence must still explain how that covariance develops.

The classic mutualism model proposed by van der Maas and colleagues starts with multiple developing cognitive processes rather than one single causal intelligence variable. Positive interactions among those processes can generate a positive manifold over time.

Imagine three abilities: efficient memory retrieval, language comprehension and reasoning. Faster retrieval can give comprehension more usable knowledge at the moment it is needed. Better comprehension can create richer and better-organised memory traces. Both can make reasoning practice more productive because relationships and rules are represented more clearly. Over years, initially distinct abilities can therefore become increasingly correlated.

The key mutualist prediction is stronger than ordinary transfer. Transfer asks whether training or growth in A produces a higher score on B. Developmental mutualism asks whether A changes the subsequent rate at which B develops.

Transfer: A improves → B is higher.

Developmental coupling: A improves → B becomes easier or faster to develop.

That distinction matters for intervention research because it shifts the target from a one-off post-test difference to a learning trajectory.

McGrew, Schneider, Decker and Bulut's 2023 psychometric network analysis provides a useful bridge between traditional CHC intelligence theory and network accounts. They analysed 3,258 people aged 9–19 using Woodcock–Johnson IV measures and represented abilities as nodes connected by conditional associations.

The resulting network was not homogeneous. Some measures were much more central than others. Three of the four most central measures — Letter–Pattern Matching, Verbal Attention and Retrieval Fluency — reflected processing speed, working-memory/attentional control and retrieval efficiency. Oral Comprehension was the most central single measure and itself draws on vocabulary, attention, inference, retrieval and working memory.

This gives a useful concrete picture. Consider a learner reading a difficult scientific paragraph. Efficient retrieval makes relevant concepts available quickly; attentional control keeps the important relation active despite distracting detail; comprehension integrates the sentence with existing knowledge. These are separable abilities, but success depends on their coordination.

The finding is compatible with a network containing local clusters, bridge processes and potentially higher-leverage nodes. It does not, however, establish causal direction. McGrew et al.'s edges were undirected partial correlations. A line between attention and reasoning cannot tell us whether attention develops reasoning, reasoning develops attention, both occur, or another process contributes to both.

That boundary is central to the theory proposed here: network structure identifies candidate dependencies; longitudinal and intervention studies must establish developmental direction.

Recent studies provide more direct evidence that some cognitive abilities can predict one another's subsequent growth.

Vocabulary and non-verbal reasoning. Griffiths, Kievit and Norbury followed 501 children aged roughly 7–13 across three waves. Models containing reciprocal coupling fitted better than uncoupled models: stronger receptive vocabulary predicted later growth in non-verbal reasoning, while stronger reasoning predicted later vocabulary growth.

A practical example is learning a new science concept. Better vocabulary can make the written explanation easier to parse; stronger reasoning can then help infer the meaning of unfamiliar terminology from the structure of the explanation. Each ability changes the learning opportunities available to the other.

Associative learning and fluid intelligence. Ren and colleagues followed elementary-school children over three years. Associative learning and fluid intelligence predicted one another at the within-person level even after working memory and processing speed were included as covariates. This is particularly relevant to mutualism because it goes beyond stable between-person correlation: doing better than expected in one process predicted later change in the other.

Fluid intelligence and need for cognition. A preregistered 2024 adolescent study found reciprocal relations between fluid intelligence and need for cognition — the tendency to seek and enjoy effortful thinking. This suggests a feedback loop between ability and the behaviour that determines how often demanding cognition is voluntarily exercised. However, the predicted cascade into crystallised intelligence was not supported.

The evidence therefore supports selected developmental couplings. It does not support the stronger idea that every cognitive ability mutually reinforces every other ability.

A four-wave 2025 study of 1,726 primary-school pupils testing mathematics achievement and cognitive reasoning found a predominantly unidirectional relationship rather than the symmetric reciprocity a simple mutualist story might suggest. Combined with the incomplete cascade in the need-for-cognition study, this points towards a more selective model.

For example, stronger attention control might make relational-memory training easier because the learner can protect the relevant relation from competing information. Relational memory might then help later reasoning because more variables and relations can be kept correctly bound. But there is no reason to assume that reasoning improves attention control by the same amount, or that every other ability joins the same loop.

A realistic developmental network might therefore contain:

• a strong Attention → Relational Memory influence;
• a weaker Relational Memory → Attention influence;
• a reciprocal Vocabulary ↔ Reasoning relationship;
• a conditional Generative Search → Reasoning link that appears mainly when tasks reward alternative generation;
• many near-zero relationships.

This is the core of Asymmetric Developmental Mutualism: developmental coupling is expected to be sparse, unequal, directional and context-dependent.

In compact form, the developmental influence of capacity i on capacity j can be represented as M_ij. The theory expects M_ij to differ from M_ji, and to vary with developmental stage and context.

Cognitive development does not happen inside an isolated brain. A 2024 review of cognitive stimulation concluded that enriching environmental inputs and learning opportunities contribute to cognitive and neural development, with intervention studies providing evidence for causal environmental effects.

Schooling provides a large-scale illustration. Quasi-experimental evidence across more than 600,000 participants suggests that additional education can increase measured cognitive ability. Sustained environments therefore do more than reveal existing ability: they can change the opportunities through which abilities are practised and extended.

But people also help create the environments they experience. In a genetically informed study of more than 15,000 young children, cognitive development and parental cognitive stimulation predicted one another across time at the observed level. Much of that association was explained by shared genetic and environmental influences rather than simple direct causation, but the study still highlights an important point: children are active participants in constructing their learning experiences.

For adults, consider two people with similar measured reasoning ability. One spends most working days following stable procedures and outsourcing difficult judgement. The other regularly compares competing explanations, learns unfamiliar systems, receives diagnostic feedback and works with colleagues who challenge assumptions. Over time, the second environment repeatedly recruits a broader set of capacities.

This motivates the idea of a cognitive niche: the recurring configuration of tasks, people, tools, information, incentives, habits and challenges through which cognition is exercised.

Suppose an analyst improves at separating relevant evidence from fluent but distracting information. That capacity may initially appear only during a training task. If it becomes strategically accessible — for example through the question "What information actually matters here?" — the analyst may start using it while reading reports, checking AI-generated summaries and comparing forecasts.

Successful use then changes behaviour. The analyst may seek more ambiguous cases, build explicit comparison steps into their workflow and ask colleagues for disconfirming evidence. The environment now recruits evidence discrimination more frequently and also places greater demands on working memory, hypothesis generation and reasoning.

An initial capacity change has therefore altered the person's future training environment. That is the multiplier idea:

capacity improves → richer or better-structured activity becomes possible → that activity repeatedly recruits several capacities → further development becomes more likely.

This does not prove that such multipliers are large in adults. It specifies a testable developmental mechanism.

Trident-G extends the developmental argument by distinguishing three intervention targets.

• Capacity: can the cognitive operation be performed?
• Strategy: is there an explicit, retrievable policy for using it?
• Niche: does the person recognise the real-world cue and repeatedly encounter an environment in which the operation is useful?

Take attention control. The capacity is the ability to protect relevant signal from interference. A portable strategy might be: "What information actually matters here?" A niche could be a research workflow in which the person routinely compares several plausible claims, records the diagnostic evidence and deliberately reduces irrelevant notifications while doing so.

Capacity without strategy may remain locked inside the training task. Strategy without capacity may collapse under load. Capacity and strategy without a supportive niche may simply not be used often enough to matter.

The stronger Trident-G hypothesis is that these three components can form feedback loops within a node and, eventually, across nodes. That proposal goes beyond the established mutualism literature and remains an intervention hypothesis.

Network neuroscience provides convergent, but not causal, evidence for a distributed account of general intelligence. A 2026 Human Connectome Project analysis found that whole-brain connectivity predicted general intelligence more effectively than any single network. Between-network communication, weak long-range connections, control properties and small-world organisation were all important features.

This does not demonstrate developmental mutualism. Neural network organisation and developmental coupling between cognitive abilities are different explanatory levels. It does, however, weaken the intuition that a psychologically general ability must map onto one privileged anatomical intelligence centre.

A system can be general because specialised components coordinate effectively.

There is an important danger in network explanations: once abilities are connected by lines, the lines start to look causal.

Huth and colleagues reanalysed 293 psychological networks from 126 papers. About one third of edges had inconclusive evidence, around half had only weak evidence, and fewer than 20% were strongly supported. The result does not invalidate network analysis, but it demands caution about individual edges.

The right hierarchy is:

conceptual network ≠ stable statistical network ≠ demonstrated causal developmental network.

McGrew et al. make a related point. Their central cognitive nodes are useful candidates for theory and intervention research, not proof that training a central node will improve the whole network.

Conventional cognitive-training research often asks:

If I train task A, does performance increase on task B?

That is important, but asymmetric developmental mutualism suggests another question:

If I develop capacity A, does capacity B subsequently become easier to develop?

For example, instead of asking only whether attention training immediately raises a reasoning score, measure Attention and Relational Memory at baseline, train Attention, and then estimate the Relational Memory learning curve. Compare that with a group that trains Relational Memory first.

If the ordering matters, the evidence is much closer to a developmental edge:

Attention → later Relational Memory learning may differ from Relational Memory → later Attention learning.

The same logic can test environmental multiplication by comparing capacity training alone with capacity plus explicit strategy and capacity plus strategy plus real-world niche deployment.

This is much stronger than assuming that a central node, a correlation or a successful brain-training game proves broad transfer. See the separate brain-training evidence guide and far-transfer guide for that claims boundary.

Well established: the positive manifold; psychometric g; meaningful broad cognitive abilities; substantial long-term stability of individual differences; environmental effects on cognitive development; and distributed neural correlates of general intelligence.

Emerging evidence: selected reciprocal developmental relationships between abilities; selected one-way relationships; links between cognitive ability and cognitively demanding behaviour; and transactional relations between people and stimulating environments.

Plausible synthesis: the developmental network is heterogeneous. Some capacities may occupy more influential positions than others, and pairwise effects need not be symmetric.

Asymmetric Developmental Mutualism: the hypothesis that selected cognitive capacities exert unequal and directional effects on the subsequent developmental rates of other capacities, with those effects varying across time and context.

Trident-G intervention hypothesis: deliberately combining capacity development, strategic accessibility and cognitive-niche redesign can create exploitable developmental multipliers in adults.

The last proposition has not yet been demonstrated. It is a research programme, not an efficacy claim.

The strongest conclusion is not that mutualism has defeated g, CHC theory or Process Overlap Theory. These frameworks operate at different explanatory levels.

• Factor models describe the structure of individual differences.
• CHC theory identifies meaningful broad cognitive abilities.
• Process-overlap accounts emphasise domain-general processes shared across tasks.
• Psychometric networks reveal conditional associations and candidate high-leverage nodes.
• Developmental mutualism asks how abilities change one another's trajectories.
• Niche-multiplier models add the environments people enter, evoke and deliberately reshape.

The resulting model is neither one mysterious intelligence substance nor a bag of unrelated skills. It is a developmental ecosystem in which capacities, strategies, learning opportunities and environments can change one another over time.

The practical target is therefore not simply "train g directly". A more defensible research aim is to identify capacities with plausible downstream leverage, make their underlying operations portable, place them into environments that repeatedly recruit them, and then measure whether those changes alter what becomes easier to learn next.

Core proposal: asymmetric cognitive synergies × niche multipliers → developmental trajectories.

What is mutualism theory of intelligence?

It is a developmental account in which distinct cognitive abilities can positively influence one another's growth. The positive manifold can therefore emerge partly from interactions among abilities rather than requiring one single causal g mechanism.

Does mutualism reject general intelligence?

No. It is compatible with using psychometric g as a useful summary of broad cognitive covariance. The disagreement concerns whether that statistical factor itself should be treated as the single causal mechanism of intelligence.

What is the positive manifold?

It is the robust tendency for scores on different cognitive tests to correlate positively. It is the empirical pattern that both latent-factor and mutualist/network theories must explain.

What makes asymmetric mutualism different?

It does not assume that every ability reinforces every other ability equally or in both directions. It predicts a sparse network containing reciprocal, mainly one-way, weak and context-dependent developmental relationships.

Does this mean brain training can raise IQ?

No. That conclusion does not follow from network centrality or developmental coupling. The stronger experimental test is whether prior development of one capacity changes the subsequent learning trajectory of another, followed by delayed and independent measurement.

1. van der Maas, H. L. J., et al. (2006). A dynamical model of general intelligence: The positive manifold of intelligence by mutualism. Psychological Review, 113, 842–861.
2. McGrew, K. S., Schneider, W. J., Decker, S. L., & Bulut, O. (2023). A Psychometric Network Analysis of CHC Intelligence Measures: Implications for Research, Theory, and Interpretation of Broad CHC Scores “Beyond g”. Journal of Intelligence, 11, 19.
3. Griffiths, S., Kievit, R. A., & Norbury, C. (2022). Mutualistic coupling of vocabulary and non-verbal reasoning in children with and without language disorder. Developmental Science, 25, e13208.
4. Ren, X., Zhao, S., Huang, X., & Lv, X. (2026). More than correlates: Longitudinal evidence of bidirectional effects between associative learning and fluid intelligence in elementary school children. Intelligence, 114, 101993.
5. Scherrer, V., Breit, M., & Preckel, F. (2024). Crystallized Intelligence, Fluid Intelligence, and Need for Cognition: Their Longitudinal Relations in Adolescence. Journal of Intelligence, 12, 104.
6. Fiedler, D., Barton, S., & Kipman, U. (2025). Cross-Lagged Relationships Between Cognitive Ability and Math Achievement. Journal of Intelligence, 13, 138.
7. Breit, M., Scherrer, V., Tucker-Drob, E. M., & Preckel, F. (2024). The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies. Psychological Bulletin, 150, 399–439.
8. Rakesh, D., et al. (2024). Environmental contributions to cognitive development: The role of cognitive stimulation. Developmental Review, 73, 101135.
9. Starr, A., Oginni, O., & von Stumm, S. (2024). Do Children Cause the Cognitive Stimulation they Receive? Modelling the Direction of Causality. Behavior Genetics, 54, 443–455.
10. Ritchie, S. J., & Tucker-Drob, E. M. (2018). How Much Does Education Improve Intelligence? A Meta-Analysis. Psychological Science, 29, 1358–1369.
11. Wilcox, R. R., et al. (2026). The network architecture of general intelligence in the human connectome. Nature Communications, 17, 2027.
12. Huth, K. B. S., et al. (2025/2026). Statistical evidence in psychological networks. Nature Human Behaviour, 10, 333–346.