STAR as Attention Architecture - A Theoretical Framework for Human-AI Cognitive Alignment

David Chadderton Creator of the STAR Framework | Chief Marketing Officer | Author

Abstract

This paper introduces the Attention Architecture Hypothesis: that the STAR Framework’s four motivational types (Socialiser, Thinker, Adventurer, Realist) can be understood not merely as psychographic categories but as distinct attention allocation strategies operating within the constraints of human cognitive architecture. By integrating Nelson Cowan’s Embedded-Process Model of working memory (1988, 2001, 2005) and John Sweller’s Cognitive Load Theory (1988), this paper proposes that STAR types differ systematically in how they allocate the limited capacity of the Focus of Attention (FoA), what schemas they construct in Long-Term Memory (LTM), and how they perform under cognitive load. The paper further argues that this integration positions STAR as a framework uniquely suited to the age of artificial intelligence, drawing structural parallels between Cowan’s three-tier memory hierarchy and transformer-based attention mechanisms. The hypothesis is presented as grounds for empirical research and as a foundation for designing AI systems that account for human attentional diversity.

Keywords: STAR Framework, cognitive architecture, working memory, attention allocation, Cognitive Load Theory, transformer architecture, psychographic segmentation, human-AI interaction

1. Introduction: Why Attention Architecture Matters Now

We are living through two simultaneous revolutions in how information is processed.

The first is human. The volume of information competing for human attention has increased by orders of magnitude in a generation, while the cognitive architecture that processes that information has not changed in millennia. The Focus of Attention, the bottleneck of conscious processing, remains limited to approximately four novel chunks of information simultaneously (Cowan, 2001). The second is artificial. Transformer-based large language models have achieved breakthrough performance not by expanding memory but by implementing sophisticated attention mechanisms that selectively weight information based on learned relevance (Vaswani et al., 2017). Both revolutions are, at their core, about the same problem: how to allocate limited attention across a vast information space.

Existing psychographic frameworks, DISC, MBTI, Insights Discovery, the Big Five, describe who people are. None of them describes how people process information. This is a significant gap. In an era where marketing, communication, education, and AI-mediated interaction all depend on matching information delivery to cognitive capacity, a framework that describes attention allocation rather than merely personality preference has structural advantages.

The STAR Framework (Chadderton, 2024) was originally developed as a motivational-psychographic model synthesising seven psychological theories: Self-Determination Theory (Deci and Ryan, 1985), the Big Five personality model (Costa and McCrae, 1992), Dual Process Theory (Kahneman, 2011), Regulatory Focus Theory (Higgins, 1997), Social Identity Theory (Tajfel and Turner, 1979), Cognitive Bias Theory, and the Appraisal Theory of Emotion (Lazarus, 1991). From these foundations, STAR identifies four core human types, each mapped to one of the four fundamental psychological needs identified by Self-Determination Theory:

The Socialiser (S) driven by Relatedness
The Thinker (T) driven by Competence
The Adventurer (A) driven by Autonomy
The Realist (R) driven by Security

Each type branches into three archetypes based on primary and secondary motivational combinations, yielding twelve distinct profiles with characteristic decision-making patterns, communication preferences, and behavioural signatures.

This paper proposes that STAR’s four types can be reinterpreted through the lens of cognitive architecture as four distinct attention allocation strategies, systematic patterns of how the Focus of Attention is deployed within the constraints of working memory. This reinterpretation has three consequences: it grounds STAR in more durable cognitive science, it generates testable predictions about information processing differences between types, and it positions STAR as a framework for designing human-AI interaction systems.

2. Cognitive Architecture: Cowan’s Embedded-Process Model

2.1 The Three-Tier Hierarchy

Nelson Cowan’s Embedded-Process Model (1988, 2001, 2005) reconceptualises working memory not as a set of separate structural buffers (as in Baddeley’s multi-component model) but as a dynamic, state-based hierarchy embedded within the brain’s existing memory networks. The model comprises three tiers:

Tier 1: Long-Term Memory (LTM). A vast network of static knowledge structures, schemas, and episodic traces. Capacity is effectively unlimited. Representations are inactive until stimulated by contextual cues or attentional control.

Tier 2: Activated Memory. A subset of LTM that has been temporarily raised above baseline activation by incoming sensory input, associative priming, or goal-directed retrieval. Activated representations are unconscious or semi-conscious. They are subject to rapid temporal decay unless refreshed by attentional control or external reactivation.

Tier 3: Focus of Attention (FoA). A capacity-limited spotlight under direct executive control. Cowan’s empirical work established that the pure capacity of the FoA is approximately N = 4 (plus or minus 1) chunks when active rehearsal strategies are blocked (Cowan, 2001). The FoA operates on activated representations, selecting which of the many currently-active memory traces enter conscious processing.

2.2 Two Attentional Systems

Cowan’s model identifies two systems that control what enters the FoA:

The Central Executive. A voluntary, goal-directed control system that directs the FoA based on internal objectives. This is effortful, capacity-limited, and responsive to current motivational state.

The Automatic Orienting System. An involuntary attentional capture mechanism driven by novel, emotionally salient, or biologically significant environmental stimuli. This system can override executive control, forcing high-salience representations into the FoA regardless of current goals.

2.3 The Capacity Constraint

The 4-chunk limit of the FoA is not a design flaw. It is a structural constraint that has been remarkably consistent across experimental paradigms, age groups, and cultural contexts. When rehearsal is prevented and chunking strategies are controlled, adults converge on approximately 3 to 4 items in the FoA (Cowan, 2001; Chen and Cowan, 2009). This limit has profound implications for instructional design, communication, and any domain where information delivery must match cognitive capacity.

3. Cognitive Load Theory: Managing the Bottleneck

3.1 Three Load Types

John Sweller’s Cognitive Load Theory (CLT) (1988, 2011) posits that learning depends on managing working memory resources to enable schema construction in Long-Term Memory. CLT identifies three types of cognitive load:

Intrinsic Load. The inherent complexity of the material being processed, determined by the number of interacting information elements that must simultaneously occupy the FoA to build a composite understanding. Intrinsic load is a function of element interactivity: a single isolated fact has low intrinsic load; a multi-variable relationship has high intrinsic load.

Extraneous Load. Cognitive capacity consumed by processing demands that do not contribute to learning or understanding. Poor formatting, split-attention layouts, decorative elements, and redundant information sources all generate extraneous load by triggering the Automatic Orienting System, diverting the FoA from task-relevant processing.

Germane Load. The productive cognitive effort dedicated within the FoA to constructing, reorganising, and integrating activated traces into LTM schemas. Germane load is the productive load: it is the mechanism by which understanding is built.

The fundamental constraint of CLT can be expressed as:

Total Cognitive Load = Intrinsic + Extraneous + Germane ≤ FoA Capacity

When total load exceeds capacity, learning fails. Schema construction is interrupted. Understanding degrades.

3.2 Schema Acquisition and Expertise

The mechanism by which expertise develops under CLT is schema construction and automation. A novice perceiving a complex problem must hold each element as a separate chunk in the FoA, rapidly exceeding the 4-chunk limit. An expert, by contrast, has constructed integrated schemas in LTM through repeated practice. These schemas enter the FoA as single chunks, dramatically reducing intrinsic load and freeing capacity for higher-order processing (Sweller and Chandler, 1994).

This has a direct implication: the same information has different cognitive load depending on the prior knowledge and schema structure of the person processing it. What is overwhelming for a novice is manageable for an expert, not because the expert has more working memory, but because they have better-chunked schemas.

4. The Attention Architecture Hypothesis

4.1 The Core Claim

The STAR Framework’s four types are not merely motivational categories. They are attention allocation strategies: systematic patterns of how the Central Executive directs the Focus of Attention within Cowan’s three-tier architecture. Each type has a characteristic attention signature, a default weighting of what gets prioritised for entry into the FoA from Activated Memory.

This claim has three components:

The FoA Allocation Claim: STAR types differ systematically in what information they prioritise in the Focus of Attention.
The Schema Construction Claim: STAR types build different types of schemas in Long-Term Memory, reflecting their motivational priorities.
The Load Sensitivity Claim: STAR types experience different sources of extraneous load, depending on the match between information delivery and their attention allocation pattern.

4.2 FoA Allocation: The Four Attention Signatures

When a STAR type encounters a new environment or information set, their Central Executive pre-weights the FoA toward information that is relevant to their core motivational need. This is not a conscious choice. It is an automatic bias in attentional allocation, shaped by years of schema construction driven by motivational priorities.

The Socialiser (S): Relational Attention. The Socialiser’s FoA is weighted toward relational cues: who is present, what the group dynamics are, where they fit in the social structure, who needs attention, and what the emotional tone of the interaction is. The Socialiser’s Central Executive generates Queries that scan for connection, inclusion, and belonging signals. When confronted with a complex information set, the Socialiser naturally chunks by relationship: this idea connects to that person; this concept relates to that group experience.

The Thinker (T): Structural Attention. The Thinker’s FoA is weighted toward logical structure: what the rules are, how the components relate, where the inconsistencies lie, and what the competence hierarchy looks like. The Thinker’s Central Executive generates Queries that scan for patterns, logical relationships, and systematic coherence. The Thinker chunks by structure: this variable interacts with that parameter; this principle governs that outcome.

The Adventurer (A): Novelty Attention. The Adventurer’s FoA is weighted toward novelty, opportunity, and change: what is new, what has changed, what possibilities exist, and what constraints can be bypassed. The Central Executive generates Queries that scan for unexplored territory and actionable momentum. The Adventurer chunks by opportunity: this opening leads to that possibility; this constraint can be circumvented.

The Realist (R): Threat Attention. The Realist’s FoA is weighted toward risk, stability, and proven patterns: what could go wrong, what is reliable, what has been tested, and what the safety margins are. The Central Executive generates Queries that scan for uncertainty and potential failure. The Realist chunks by reliability: this approach has a track record; that option introduces unacceptable risk.

4.3 Schema Construction: Types of Expertise

If STAR types allocate attention differently, they will also construct different types of schemas in LTM over time. This is not merely a motivational difference. It is a structural difference in the knowledge representations that each type builds.

Socialisers develop dense relational schemas: maps of social networks, emotional dynamics, group norms, and interpersonal patterns. These schemas are highly interconnected, with many associative links between social representations.

Thinkers develop dense analytical schemas: hierarchical knowledge structures organised by logical relationships, causal chains, and systematic principles. These schemas are deep and well-structured, with clear inferential pathways.

Adventurers develop dense procedural schemas: action-oriented knowledge structures organised by sequences, possibilities, and outcomes. These schemas are flexible and cross-contextual, with many lateral connections between different domains of experience.

Realists develop dense verification schemas: knowledge structures organised by evidence, precedent, and reliability. These schemas are conservative and well-tested, with strong connections to empirical evidence and past experience.

These different schema structures have direct implications for expertise development. A Thinker’s analytical schema may enable them to chunk complex logical relationships into a single FoA unit, but may not help them process relational information efficiently. A Socialiser’s relational schema may enable them to instantly read group dynamics, but may not help them process technical specifications. Each type’s schema structure is an expertise architecture that enables certain kinds of processing while providing no advantage for others.

4.4 Load Sensitivity: Differential Extraneous Load

If STAR types have different attention allocation patterns, then information delivery formats that match one type’s pattern will generate germane load for that type but extraneous load for another. This is a testable prediction.

Consider a marketing message that leads with social proof (Join 50,000 happy customers). For the Socialiser, this is germane load: it activates relational schemas and directs the FoA toward belonging-relevant information. For the Thinker, this is extraneous load: it consumes FoA capacity without providing the structural information the Thinker’s Central Executive is seeking.

Conversely, a message that leads with detailed specifications generates germane load for the Thinker but extraneous load for the Socialiser, whose Central Executive is scanning for relational cues that the specifications do not provide.

This differential load sensitivity predicts that the same information, delivered in the same format, will have different cognitive costs for different STAR types. This is not a preference difference. It is an attentional architecture difference. And it has direct implications for how content, communication, and instructional design should be structured.

5. The Transformer Parallel: Attention in Biological and Artificial Systems

5.1 The Structural Isomorphism

The transformer architecture (Vaswani et al., 2017), which underlies modern large language models, solves a problem that is structurally identical to the one Cowan’s model describes. Both systems must allocate limited attention across a vast information space. Both achieve this through a selective attention mechanism that weights different information elements differently based on relevance.

The mapping is as follows:

Cowan’s Long-Term Memory corresponds to model weights in a transformer: pre-trained knowledge that is static until activated. Activated Memory corresponds to the context window: currently active tokens, subject to length limits. The Focus of Attention corresponds to the self-attention mechanism: selectively weighted, capacity-constrained. The Central Executive corresponds to Query vectors: goal-directed attention allocation. The Automatic Orienting System corresponds to high-attention tokens: novelty, salience, and positional signals that capture processing resources.

The transformer’s self-attention mechanism computes attention weights by comparing Queries (what the model is currently trying to process) against Keys (what information is available in the context window). When a Query matches a Key, the corresponding Value (content) receives high attention weight and contributes to the output. This Query/Key/Value mechanism is functionally equivalent to Cowan’s Central Executive (Query) selecting from Activated Memory (Keys) to determine what enters the FoA (Values).

5.2 The Attention Bottleneck in Both Systems

Both systems face the same fundamental constraint. The transformer’s context window, like Cowan’s FoA, has a fixed capacity. When the context window is full, earlier information must be pruned, compressed, or moved to external storage. When the FoA is saturated, additional information cannot be processed.

The solutions are also parallel. Transformers use techniques like sparse attention, sliding windows, and hierarchical attention to manage context limits. Humans use chunking, schema automation, and external memory aids to manage FoA limits. Both systems are optimising the same trade-off: breadth of attention versus depth of processing.

5.3 STAR Types as Attention Head Configurations

In multi-head transformer architectures, different attention heads learn to attend to different types of information: some heads track syntax, others track semantics, others track positional relationships. The diversity of attention heads is what gives transformers their representational power.

STAR’s four types can be understood as analogous to differently configured attention heads. Each type has learned, through a lifetime of motivational prioritisation, to weight different aspects of incoming information. The Socialiser has a relational head that weights interpersonal cues. The Thinker has a structural head that weights logical relationships. The Adventurer has a novelty head that weights change and opportunity. The Realist has a threat head that weights risk and uncertainty.

This is not merely an analogy. If human attention allocation follows the same computational principles as transformer attention, then STAR’s types are descriptions of optimised attention configurations, solutions to the problem of how to allocate limited cognitive resources given different motivational objectives.

5.4 Implications for AI Design

If STAR types represent distinct attention allocation strategies, and if transformer architectures implement analogous attention mechanisms, then there are direct implications for AI system design:

Personalised AI Attention. Current AI systems apply the same attention weights to all users. If users differ in their attention allocation patterns (as STAR predicts), then AI systems could adapt their information delivery to match the user’s type. A STAR-aware AI presenting information to a Socialiser would lead with relational framing; presenting the same information to a Thinker, it would lead with structural framing. The underlying content is identical. The attention allocation is different.

AI-Mediated Discovery. In an AI-mediated discovery environment (such as AI-powered search, recommendation systems, or conversational agents), the AI is effectively performing attention allocation on behalf of the user. The AI selects what information to surface, in what order, with what framing. If STAR describes human attention patterns, it can also describe the optimal shape of AI-mediated information delivery for different user types.

Human-AI Cognitive Alignment. The deepest implication is that AI systems could be designed to align with human cognitive architecture, not just human preferences. Current AI optimisation focuses on what users want. STAR-informed AI would also account for how users process. This is the difference between delivering the right content and delivering the right content in the right attentional configuration.

6. The Stress-Performance Nexus: A Mechanistic Account

6.1 Anxiety and FoA Capacity

Cowan’s model provides a mechanistic explanation for the well-documented decline in cognitive performance under stress. When an individual experiences performance anxiety or physiological stress, threat-related thoughts (fear of failure, social evaluation, uncertainty) function as high-salience representations. Due to evolutionary priority, the Automatic Orienting System forces these representations into the FoA, regardless of the Central Executive’s current goals.

If intrusive anxious thoughts consume 2 of the FoA’s 4 available chunks, effective capacity drops to 2 chunks. This is not a gradual degradation. It is a step-function reduction in cognitive bandwidth that explains the sharp decline in problem-solving ability, decision quality, and skilled performance observed under high-stakes conditions (Eysenck et al., 2007).

6.2 Differential Anxiety Triggers by STAR Type

The Attention Architecture Hypothesis predicts that STAR types will be differentially vulnerable to cognitive degradation under stress, because their anxiety triggers are rooted in different motivational needs:

Socialisers experience cognitive load reduction when excluded, socially evaluated, or when belonging is threatened. The Automatic Orienting System captures on social rejection cues.

Thinkers experience cognitive load reduction when their competence is challenged, when they encounter logical inconsistency without resolution, or when they lack the knowledge to understand a situation. The system captures on incompetence signals.

Adventurers experience cognitive load reduction when their autonomy is constrained, when they are micromanaged, or when forced into rigid structures without flexibility. The system captures on constraint signals.

Realists experience cognitive load reduction when facing uncertainty, untested situations, or when reliable structures are disrupted without clear alternatives. The system captures on instability signals.

This differential vulnerability has practical implications for team composition, leadership, and performance coaching. A team composed entirely of Adventurers will degrade under constraint pressure. A team composed entirely of Realists will degrade under change pressure. Optimal team design accounts for complementary stress vulnerabilities.

6.3 Schema Robustness as Stress Inoculation

The military training paradigm, particularly in high-stakes aviation environments, provides empirical support for the schema-robustness hypothesis. Military instructors do not train personnel to manage stress in a general sense. They train them to build schemas so robust and so well-practised that they can function effectively on 2 FoA chunks instead of 4. This is achieved through:

Overlearning: Repeating procedures until they are automated and consume minimal FoA capacity.

Stress inoculation: Practising under controlled stress conditions to build schemas that remain accessible when the Automatic Orienting System activates.

Chunking optimisation: Structuring procedures into 3-4 step sequences that align with FoA capacity limits.

This training paradigm is, in Cowan’s terms, a systematic programme for building schemas that are resistant to capacity reduction under stress. It is the empirical foundation for the claim that attention architecture can be deliberately shaped.

7. Research Programme: Testing the Hypothesis

7.1 Empirical Predictions

The Attention Architecture Hypothesis generates several testable predictions:

Prediction 1: Differential attention allocation. When presented with identical complex information environments (e.g., a website, a presentation, a product page), participants classified by STAR type will show different eye-tracking patterns, different recall of specific information elements, and different information-seeking sequences. Socialisers will attend first to social cues (reviews, testimonials, human images). Thinkers will attend first to structural information (specifications, comparisons, logical arguments). Adventurers will attend first to novelty signals (new features, unique claims, change indicators). Realists will attend first to reliability signals (guarantees, track records, risk disclosures).

Prediction 2: Differential cognitive load. The same information format will generate different subjective cognitive load ratings (measured by NASA-TLX or equivalent) for different STAR types, depending on the match between the format and the type’s attention allocation pattern. Information formatted for one type will be experienced as higher load by other types.

Prediction 3: Differential schema construction. After learning identical material, participants will show different schema structures in recall tasks, concept mapping, and transfer tests. Socialisers will organise knowledge by relational themes. Thinkers will organise by logical structure. Adventurers will organise by application and possibility. Realists will organise by evidence and reliability.

Prediction 4: Differential stress vulnerability. Under induced cognitive stress (e.g., time pressure, ego threat, uncertainty), different STAR types will show performance degradation in tasks that tap their specific vulnerability. Socialisers will degrade under social exclusion. Thinkers will degrade under competence challenge. Adventurers will degrade under autonomy constraint. Realists will degrade under uncertainty.

Prediction 5: Transformer attention alignment. When STAR-typed users interact with AI systems, the AI’s internal attention patterns (measurable via attention weight analysis) will show partial alignment with the user’s STAR type. Users who are Socialisers will generate prompts and queries that produce higher attention weights on relational tokens. This prediction is testable using interpretability methods on transformer-based models.

7.2 Methodology

Testing these predictions requires a mixed-methods approach:

Phase 1: Classification validation. Confirm that STAR type classification (using the existing STAR CPA instrument) predicts measurable differences in attention allocation using eye-tracking, information search behaviour, and recall protocols.

Phase 2: Load measurement. Use dual-task paradigms and subjective load ratings to measure differential cognitive load across STAR types for identical information formats.

Phase 3: Schema analysis. Use concept mapping, free recall, and transfer tasks to characterise the schema structures that different STAR types construct from identical learning materials.

Phase 4: Stress manipulation. Use established stress induction paradigms (Trier Social Stress Test, time pressure, ego threat) to test differential vulnerability predictions.

Phase 5: AI alignment analysis. Use attention weight analysis and user studies to test whether AI systems’ attention patterns partially align with users’ STAR types.

8. Implications

8.1 For Marketing and Consumer Psychology

If STAR types have different attention allocation patterns, then marketing content should be structured to match. This goes beyond personalisation in the current sense (which typically means adjusting offers or recommendations). It means adjusting the attentional entry point of information:

For Socialisers: lead with relational framing (community, belonging, social proof). For Thinkers: lead with structural framing (data, logic, systematic comparison). For Adventurers: lead with novelty framing (opportunity, possibility, freedom). For Realists: lead with reliability framing (evidence, precedent, guarantees).

The underlying product information is identical. The attentional architecture of the delivery is different.

8.2 For AI-Mediated Discovery

The Invisible First Click framework (Chadderton, 2025) argues that in an AI-mediated environment, the consideration set is formed before the consumer ever visits a website. The AI has already made recommendations, established credibility hierarchies, and effectively pre-selected a shortlist. STAR’s attention architecture layer provides a framework for understanding how AI systems should present information to different user types. If the AI knows (or infers) the user’s STAR type, it can structure its outputs to match their attention allocation pattern, reducing cognitive load and increasing the probability that the information will be processed, understood, and acted upon.

8.3 For Leadership and Team Design

The differential stress vulnerability predictions have direct implications for team composition. The DCTA framework (Drivers, Custodians, Translators, Aligners) already describes team kinetic profiles. The attention architecture layer adds a cognitive dimension: teams need complementary attention allocations and complementary stress vulnerabilities. A team where all members degrade under the same stressor is structurally fragile. A team where members’ vulnerabilities are distributed across different stressors is resilient.

8.4 For AI Architecture

The deepest implication is for how AI systems are designed. Current AI systems are type-agnostic: they apply the same processing to all users. If STAR’s attention signatures are real, then AI systems could be designed with type-aware attention mechanisms that adapt their information processing to match the user’s cognitive architecture. This is not personalisation in the recommendation-engine sense. It is cognitive alignment: matching the AI’s attention allocation to the human’s attention allocation.

9. Conclusion: From Psychography to Cognitive Architecture

The STAR Framework was built on seven psychological pillars to describe what drives people. This paper proposes an eighth dimension: how people allocate attention. By integrating Cowan’s Embedded-Process Model and Cognitive Load Theory, STAR gains a cognitive architecture layer that is more resistant to theoretical revision than its psychological foundations. The 4-chunk limit of the Focus of Attention is an empirical constant. The three-tier memory hierarchy is a structural fact. These are closer to engineering constraints than psychological theories.

The transformer parallel is not metaphor. It is structural isomorphism. Both human cognition and transformer architecture solve the same computational problem: allocating limited attention across a vast information space. If STAR describes how humans solve this problem differently based on motivational priorities, then STAR is not merely a psychographic framework. It is an attention architecture framework that describes how different minds process information, how they construct expertise, how they degrade under stress, and how AI systems should be designed to work with them.

No existing psychographic framework, DISC, MBTI, Insights Discovery, the Big Five, has a cognitive architecture component. None describes how its types allocate attention. None generates predictions about schema construction, cognitive load, or stress vulnerability. None draws structural parallels to the computational architecture of modern AI. STAR, with the attention architecture layer, is the first to do so.

The hypothesis presented here is empirically testable. The research programme is feasible. The implications for marketing, AI, leadership, and human-computer interaction are significant. The grounds for integration into AI system design are structurally sound.

STAR began as a framework for understanding what motivates people. It is becoming a framework for understanding how people think.

References

Chadderton, D. (2024). The STAR Framework: Rewriting the Rules of Consumer Engagement. NYC Big Book Award 2025.

Chadderton, D. (2025). The STAR Operating System: Decode Mindset, Understand Motivation, Transform Human Behaviour.

Chen, Z., and Cowan, N. (2009). Chunk limits and length limits in immediate recall: A reconciliation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(6), 1498-1506.

Costa, P. T., and McCrae, R. R. (1992). Revised NEO Personality Inventory (NEO-PI-R) and NEO Five-Factor Inventory (NEO-FFI) professional manual. Psychological Assessment Resources.

Cowan, N. (1988). Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system. Psychological Bulletin, 104(2), 163-191.

Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87-114.

Cowan, N. (2005). Working memory capacity. Psychology Press.

Deci, E. L., and Ryan, R. M. (1985). Intrinsic motivation and self-determination in human behavior. Plenum Press.

Eysenck, M. W., Derakshan, N., Santos, R., and Calvo, M. G. (2007). Anxiety and cognitive performance: Attentional control theory. Emotion, 7(2), 336-353.

Higgins, E. T. (1997). Beyond pleasure and pain. American Psychologist, 52(12), 1280-1300.

Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.

Lazarus, R. S. (1991). Emotion and adaptation. Oxford University Press.

Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257-285.

Sweller, J., and Chandler, P. (1994). Why some material is difficult to learn. Cognition and Instruction, 12(3), 185-233.

Sweller, J., Ayres, P., and Kalyuga, S. (2011). Cognitive load theory. Springer.

Tajfel, H., and Turner, J. C. (1979). An integrative theory of intergroup conflict. In W. G. Austin and S. Worchel (Eds.), The social psychology of intergroup relations (pp. 33-47). Brooks/Cole.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.