Intelligence as Relational Structure: Cognition, Graphs, Projection, and Quantum Models

Executive Summary

Intelligence is best treated as a real but non-substantial pattern: a stable, causally effective organization of relationships among cognitive abilities, brain networks, bodies, tools, and other minds. Several independent research traditions point in this direction. Psychometrics models intelligence as a general factor derived from correlations among tasks.[1] Network accounts explain those correlations through reciprocal interactions among abilities.[2] Network neuroscience locates general intelligence in the dynamics of distributed brain networks.[3][4] Embodied, extended, and enactive theories place cognition in loops that include the body, tools, and other people.[5][6][7]

This paper traces that argument in stages. It separates intelligence from cognitive ability, reviews what is inherited and what is learned, asks whether intelligence is a perception or a capability, examines the claim that relations matter more than nodes, and considers artificial intelligence. It then examines serialization and projection as bottlenecks in both human and machine cognition, and closes with two quantum-related threads: quantum computing on graphs and quantum-like models of cognition. Evidence strength differs sharply across these sections, and each is labeled accordingly.

Defining the Terms

Intelligence

Intelligence is usually defined as a broad mental capacity for reasoning, problem solving, learning, and adaptation across many kinds of tasks.[8] Psychometric research models this as a general factor, g, representing shared variance across diverse cognitive tests.[8] The g factor is a dimension in the statistical sense: it is extracted from correlations among tasks, its value depends on which tasks are included, and its causes remain debated.[1]

Cognitive ability

Cognitive abilities are narrower capacities defined by success on particular classes of tasks, such as working memory, processing speed, verbal comprehension, or spatial reasoning.[8] Real tasks usually draw on several at once, so test scores reflect combinations of abilities.[8][9]

Overlap

The two concepts overlap most where specific processes scale into general performance. Working memory, attention control, inhibition, and relational integration are strong candidates.[8][10][9] Relational integration, the ability to jointly consider multiple relations to reach a higher-order inference, is a strong predictor of complex cognitive performance and helps explain fluid reasoning.[11][10][12] Intelligence can therefore be described as the system-level capability, with cognitive abilities as its contributing functions.

Inherited and Learned Components

Heritability is a population statistic describing how much variation among people in a given environment is associated with genetic differences. It does not describe how much of an individual's intelligence is genetic or fixed.[13][14] Estimates for general cognitive ability rise across development; one large twin analysis reported 41% in childhood, 55% in adolescence, and 66% in young adulthood.[15][16] Cognitive ability is highly polygenic, with many variants of small effect.[13][17]

Genes and environment interact. Active gene–environment correlation, in which people select and create environments aligned with their predispositions, is one proposed explanation for rising heritability across the life course.[18] Early environments also matter; longitudinal research links supportive early environments to later cognitive ability.[19]

Change in adulthood

Adult cognition changes in uneven ways. Crystallized abilities, such as vocabulary and domain knowledge, tend to remain stable or grow, while fluid abilities and processing speed tend to decline on average with age.[20][21] Exercise trials in older adults report modest gains in attention, processing speed, executive function, and memory, and combined cognitive and physical training has also been studied.[22][23] The WHO guidelines on reducing cognitive decline emphasize physical activity and other modifiable risk factors.[24][25]

The practical conclusion is that functional capability, meaning expertise, strategy, judgment, and the use of tools, can keep improving, while large gains in a global score are less well supported.

Is Intelligence a Perception?

Two questions are easily conflated. The epistemic question asks how intelligence is recognized; the Turing Test takes a behavioral route, treating indistinguishable conversational performance as evidence of intelligence.[26] The ontological question asks what intelligence is; one account defines it as the capacity of an agent that takes in percepts and acts effectively toward goals.[27][28]

Intelligence is therefore not merely a matter of how impressive something appears. It is also relational: it is expressed relative to a task, environment, objective, and set of constraints. Observation is evidence about intelligence, not a replacement for it.

The individual and the shared

Perception and deliberation occur inside one mind, but their value is tested when decisions change a shared world and feedback returns. Enactivism supports this loop view; it describes cognition as an ongoing coupling between agent and world, and extends sense-making into the social domain as participatory sense-making.[29][7][30] The extended mind thesis holds that when external elements play an active causal role in cognition, person and element form a coupled cognitive system.[6][31] Socially extended accounts apply the same logic to shared intentionality.[32]

Collective intelligence research provides an empirical anchor. A general collective intelligence factor explains group performance across tasks, and it is not strongly correlated with the average or maximum individual intelligence of members; it correlates with average social sensitivity and with equality of conversational turn-taking.[33][34][35] Individual capability becomes impact only through coordination.

Edges Before Nodes: Intelligence as Graph

Graph-like structure, not visual form

Intelligence is not inherently visual. It is plausibly relational and network-like. Reasoning tasks require representing entities and the relations among them, and the difficulty of such tasks rises with the number of relations that must be held together.[10][9][36] Network neuroscience describes general intelligence as emerging from the flexibility, efficiency, and dynamic reconfiguration of connections across brain networks.[3][4]

Mutualism

The mutualism model offers a concrete case of edges taking priority. It shows that the positive manifold can emerge purely from beneficial interactions among cognitive processes during development, without positing a single underlying g in the model.[2] In this account, abilities are nodes, their reciprocal influences are edges, and global intelligence is a property of the whole pattern.[37][38] The model is an alternative explanation, and the field has not settled the question.[39]

Ontological status

Philosophers distinguish realism, nominalism, and conceptualism about universals.[40] The g factor illustrates the dispute: critics such as Stephen Jay Gould argued that g is a reified construct, while defenders argue that treating extracted factors as candidate causes is ordinary scientific practice.[1]

A relational ontology has a counterpart in ontic structural realism, which in its strong form treats relations as more fundamental than objects.[41][42] The standard objection is that relations seem to require relata.[43][42] A defensible middle position is pattern realism: a pattern, like a whirlpool in water, can be real without being a separate substance. On that reading, intelligence is a property of graph dynamics, how well a network integrates information, adapts, and acts, while nodes still constrain what edges can do.

Artificial Intelligence

AI is commonly defined as the field of building agents that perceive and act effectively, and under that functional definition contemporary systems display real competence.[27][44] Whether they are intelligent in the fullest human sense depends on the criterion used.

Position	Core claim	Implication for AI
Behavioral	Reliable performance of the relevant functions is intelligence	Capable AI is intelligent[26]
Rational agent	Intelligence is effective goal-directed action under uncertainty	Artificial agents qualify[27][28]
Understanding-first	Semantic understanding is required	Behavior may simulate understanding[45]
Consciousness-first	Subjective experience is required	Competence may exist without it[46]
Embodied	Intelligence is grounded in a body acting in a world	Text-only systems are limited[5]

Intelligence and consciousness can be dissociated, since AI systems appear to exhibit intelligence without consciousness.[46] A useful distinction separates capability (intelligence), experience (consciousness), sustained goal pursuit (agency), and value-sensitive judgment (wisdom). Current systems have clear evidence for the first and limited, externally scaffolded forms of the third.

Serialization and Projection

Projection theories

Projection theories hold that cognition and data are lossy, lower-dimensional views of a richer reality. One formulation describes reducing reality into lawful, lossy representations and treats knowledge and prediction as properties of the projection kernel.[47] Another proposes that cognition is the processing of low-dimensional projections.[48] Because several frameworks share this name, the term is used here in this general sense.

The Platonic Representation Hypothesis fits the picture: it argues that different modalities are projections of an underlying reality and that networks trained on different data and objectives converge toward a shared statistical model of it.[49][50] The hypothesis is a position paper and has limitations and counterexamples.[49]

Serialization as a bottleneck

Language models learn through token sequences, which flatten graph-structured information into a line. Autoregressive generation is strictly sequential and non-erasable, which complicates backtracking and revision.[51] Reviews report that language models handle single triplets well but struggle with structural reasoning, and list limitations such as context window limits, structural hallucination, and permutation sensitivity.[52][53] Permutation sensitivity is a direct serialization problem: the same graph listed in a different order can change behavior.

Some architectures attempt to undo the flattening. The Graph Transformer Language Model resets positional encodings per node, uses a fully visible attention mask over the graph prefix, and computes structural biases from graph topology rather than sequence position, with a reported quadratic attention cost.[54] Retrieval from external graph stores is another route, and fits the extended mind view in which an external structure is part of the cognitive system.[6]

Humans serialize too, since speech and writing are one-dimensional, so serialization is a shared interface and not uniquely artificial. What differs is the architecture's handling of structure on either side of it.

Quantum Threads

Measurement as projection

Quantum theory contains a precise projection operation. Under the projective formulation of the Born rule, outcome probabilities are given by overlaps of the state with projections onto measured subspaces.[55] Measurement is non-unitary and discards information, and the collapse postulate remains controversial.[55][56] This is a suggestive analogy for lossy projection in cognition, but no evidence found here links the two causally.

Quantum computing on graphs

Quantum walks give a quadratic speedup for finding a marked node, similar to Grover's algorithm.[57] Generic walk-based speedups known so far are at most quadratic, and advantage depends on graph structure; one study trained a network to predict quantum speedup on graphs of up to 25 vertices.[58][59] Quantum walks have also been proposed as an approach to quantum neural networks, with modest speedups in some regimes.[60]

The input problem

Loading classical data into quantum states is itself a bottleneck. A recent preprint argues that the worst-case cost is Θ(N), as a counting argument and not a hardware limitation, and that this undermines advantage for quantum machine learning on classical data.[61] It suggests that advantage is more plausible where states are efficiently preparable, device-generated, or variationally learned.[61] This is a single recent preprint and should be treated as an argument under test.

Quantum consciousness

Orchestrated objective reduction proposes that consciousness arises from quantum events in neuronal microtubules.[62][63] Critics argue that the brain is too warm, wet, and noisy to maintain coherence, citing Tegmark's decoherence estimates.[64][65] A 2026 review concludes that no experimental data currently shows quantum coherence in neural microtubules in vivo or its necessity for consciousness.[65] The hypothesis concerns consciousness, which is distinct from intelligence.

Quantum-Like Cognition

Quantum cognition borrows the mathematics of quantum probability, including Hilbert spaces, superposition, projection, and contextuality, to model judgment and decision without assuming physical quantum processes in the brain.[66][67][68]

Formal concept	Cognitive reading
State vector	Belief state over a set of questions[69]
Superposition	Deep uncertainty or ambivalence[67]
Projection	Answering a question or making a judgment[70]
Non-commuting projections	Question order effects[70][71]
Interference	Deviations from classical probability[72]
Entanglement	Non-decomposable meaning in concept combinations[68]

In this formalism, answering a question projects the state onto that question's subspace and changes it, so a second question starts from an altered state. This yields order effects, where the probability of A then B differs from B then A.[70][71] Order is therefore not merely formatting: a sequence of non-commuting projections changes outcomes, which is a precise sense in which serialization matters.

The same mathematics has been proposed for language model semantics. One proposal treats the embedding space as a complex Hilbert space, doubling dimensionality to provide degrees of freedom for interference and superposition of semantic states.[73] Others encode words as static vectors in a Hilbert space using quantum contextuality, or extend the idea to N-dimensional spaces with dynamic Hamiltonians and decoherence.[74][75] These are largely preprints and proposals.

Limits are explicit in the literature. A Hilbert-space structure in behavior does not imply physical quantum processes in neural tissue, and the models' explanatory value depends on empirical calibration.[76] Most approaches that apply quantum theory to brain states are described as speculative, and quantum-like models differ on the source of order effects.[72][70]

Synthesis

A coherent framework follows from these threads.

Nodes are abilities, brain regions, people, tools, and concepts.
Edges are reciprocal influences, integrations, and feedback loops among nodes.
Intelligence is a property of the dynamics of the resulting graph: integration, adaptation, and effective action.
Impact arises when this graph couples to other minds, tools, and environments, and the resulting feedback calibrates the system.
Serialization and projection describe the interfaces through which structure is flattened, transmitted, and reconstructed.
Quantum formalism offers a candidate mathematics for context-dependent judgment and for graph algorithms, with evidence of differing strength.

Evidence tiers

Claim	Status
Heritability rises across development; environment and learning matter[15][19]	Well supported
Intelligence correlates with network-level brain organization[4]	Supported, active research
Mutualism explains the positive manifold[2]	Credible alternative, contested[39]
Relational integration underlies fluid reasoning[10][11]	Supported
Collective intelligence factor in groups[33]	Supported by multiple studies[34]
Embodied and extended cognition[5][6]	Influential, partly contested
Platonic convergence of representations[49]	Hypothesis with limitations
Quantum-walk speedups on graphs[57][59]	Established but limited in scope
Quantum input bottleneck[61]	Recent preprint
Quantum-like models of judgment[70][76]	Active, behavior-level only
Quantum effects underlying consciousness[65]	Not supported by direct evidence

Open Questions and Tests

Is the serial bottleneck a limit on intelligence or on current architectures? A comparison of agent performance when reasoning over a serialized graph versus querying a graph store would measure where structure is lost.
Do language models show human-like order effects? Presenting paired questions in both orders and testing for non-commutativity would probe whether quantum-like structure describes model behavior.
Can a knowledge graph be represented natively as a Hilbert-space model? Modeling states as belief vectors and questions as projections would test whether the formalism reproduces order effects that a classical graph model does not.
Which graphs benefit from quantum walks? Predicting speedup from structure, as in the 25-vertex study, could be extended to real knowledge graphs.[58]
Where do edges fail to explain? If abilities could be shown to be independent modules whose performance never depends on interaction, the edge-first view would weaken. Development and network evidence so far points the other way.[2][4]

Caveats

Nodes are not empty; capacity limits, such as working memory, constrain what relations can achieve.
Describing intelligence as a graph is a modeling choice that fits well, not proof of its ultimate nature.
Describing intelligence as a dimension is a statistical statement until a deeper basis is shown.[1]
Impact is evidence about intelligence over many decisions, not a definition of it; luck, power, and access also shape outcomes.
Analogies among projection in cognition, serialization in language models, and quantum measurement are suggestive and should not be treated as mechanisms.