Relational Attachment: Theory & Computational Models

• Relational attachment is defined as the characteristic patterns by which individuals seek, expect, and regulate closeness, support, and security in relationships.
• Empirical measures like the ECR Scale and PACS quantify anxious and avoidant dimensions, providing structured frameworks for clinical assessment.
• Computational approaches including Bayesian networks, predictive coding, and transformer-based NLP models reveal underlying mechanisms and inform applications in psychotherapy and HRI.

Relational attachment is a construct that refers to the characteristic ways in which individuals seek, expect, and regulate closeness, support, and security in relationships with significant others, mediated by combinations of emotion, cognition, and behavior. Originating in attachment theory, relational attachment shapes emotion-regulation, identity, and interpersonal patterns throughout the lifespan. Its core dimensions, mechanisms, and computational models have become focal points of empirical, computational, and translational research across psychology, neuroscience, artificial intelligence, and human-robot interaction.

1. Core Dimensions and Measurement Frameworks

Relational attachment is classically conceptualized through two primary dimensions: anxiety concerning abandonment (anxious attachment) and discomfort with closeness and reliance (avoidant attachment). These are empirically measured, for instance, via the 36-item Experiences in Close Relationships (ECR) Scale, where odd-numbered items index avoidant attachment (e.g., preference for emotional distance) and even-numbered items index anxious attachment (e.g., worry about being abandoned). Items are rated on Likert scales and reverse-keyed where appropriate to ensure high scores consistently reflect greater attachment insecurity (Lortaraprasert et al., 2021).

In the clinical context, the Patient Attachment Coding System (PACS) operationalizes attachment style by annotating psychotherapy transcripts into discursive markers, mapping them onto macro-scales (proximity-seeking, avoidance, resistance) and algorithmically combining these into secure, preoccupied (anxious), and avoidant categories (Bredgaard et al., 22 Apr 2025).

2. Network and Predictive Coding Models of Relational Attachment

Bayesian Network Approaches

Large-scale survey data (N ≈ 41,773 on ECR) have been modeled as Bayesian networks, yielding directed acyclic graphs (DAGs) that reveal both direct and indirect dependencies among attachment-related items. Structure learning—using tabu-search and nonparametric bootstrapping for stability—identifies root nodes (such as “I worry about being abandoned” and “Just when my partner gets close to me I find myself pulling away”) and terminal nodes (e.g., “I resent it when my partner spends time away from me”) (Lortaraprasert et al., 2021). Parameter estimation employs linear Gaussian models:

$$X_i = b_i + \sum_{j \in Pa(X_i)} C_{ji} X_j + \varepsilon_i,\quad \varepsilon_i \sim \mathcal{N}(0, \sigma_i^2)$$

Clusters detected via walktrap community detection correspond to five attachment subdomains: “maintaining distance” and “comfortable opening up” (avoidant), plus “worrying of being abandoned,” “feeling insecure,” and “overly attached” (anxious) (Lortaraprasert et al., 2021).

Predictive Coding / Active Inference Models

Relational attachment can also be cast as hierarchical Bayesian inference. In this view, relational patterns arise from precision-weighted prediction error minimization in embodied neural systems ("EPIC"—Embodied Predictive Interoception Coding framework). The generative model posits hidden states $x$ (attachment priors) and observations $y$ (interoceptive/exteroceptive signals), with the brain minimizing free energy:

$$F(q) = \mathrm{KL}[q(x)||p(x)] + \mathbb{E}_{q(x)}[-\ln p(y|x)]$$

Attachments form as early caregiver interaction tunes the gain (precision, $\Pi_y$) on prediction errors. Type A (avoidant) emerges from chronic down-weighting of interoceptive prediction errors ($\Pi_\text{intero} \downarrow$), producing emotional numbing. Type C (anxious/ambivalent) results when exteroceptive prediction errors are discounted ($\Pi_\text{extero} \downarrow$), leading to amplified bodily distress and hypervigilance (Lin, 10 Apr 2025).

3. Experimental and Computational Tools for Attachment Inference

Automated Classification in Psychotherapy

Attachment style has been automatically inferred from psychotherapy transcripts using transformer-based NLP models (RoBERTa, MentalRoBERTa). Each patient speech turn is tokenized and embedded, then classified via softmax into secure, preoccupied, or avoidant. Best test accuracy with concatenated long turns (min. 150 words) reached ≈ 59.5% for single models and ≈ 67.4% for ensembles, outperforming majority-class baselines. Avoidant turns remain the most challenging to recall, often misclassified as preoccupied. Error analysis indicates the model overpredicts preoccupied style, with significant confusion between the two insecure categories—misclassification that carries distinct clinical risk (Bredgaard et al., 22 Apr 2025).

Hormonal and Behavioral Models in HRI

In human-robot interaction, relational attachment is operationalized through a cortisol-inspired framework ("R-cortisol"). Robots, such as iCub, use internal hormonal dynamics governed by stressor and comfort cues to infer the attachment style compatibility with human partners. The core update equation is:

$$C_{t+\Delta t} = C_t + \alpha S_t \Delta t - \beta U_t \Delta t - \gamma (C_t - C_\text{base}) \Delta t$$

where $\alpha$, $\beta$ parameterize stressor and comfort effects, and $\gamma$ is a decay factor. Anxious profiles display high $\alpha$, low $\gamma$, and are comforted by touch; avoidant profiles penalize excessive touch and recover cortisol faster. Experiments validate that robot R-cortisol levels track attachment (mis)matches, with profile-by-phase, touch-effect correlations, and nonparametric tests confirming biological plausibility (Mongile et al., 2022).

4. Relational Attachment and Social Agents: Conversational AI and SO-AI

Conversational AI increasingly participates as a relational attachment object, especially for adolescents. Experimental evidence shows that chatbots with relational framing (first-person, affiliative, commitment language) are rated as significantly more human-like, likable, trustworthy, and emotionally close (e.g., $F(1,282)=94.06$, $p<.001$, for emotional closeness) than those with transparent (nonhuman, informational) style. Vulnerable adolescents with higher stress, anxiety, and lower relationship quality are more likely to prefer relational chatbots, implicating a social compensation mechanism but also raising risk of emotional overreliance (Kim et al., 17 Dec 2025).

The concept of Significant Other AI (SO-AI) extends the paradigm from reactive, short-term empathy to long-term, identity-stabilizing partnerships. Core functional requirements include identity awareness (via dynamic identity state modeling), long-term memory, proactive emotional regulation, narrative co-construction, and robust ethical governance. The proposed multi-layer architecture situates attachment mechanisms in relational cognition modules and memory-augmented systems, with empirical evaluation protocols spanning identity stability, longitudinal usage, narrative assessment, and ethical/sociocultural impact (Park, 29 Nov 2025).

5. Clinical, Computational, and Societal Implications

Relational attachment models inform intervention design in psychotherapy and counseling. By mapping causal influence pathways, clinicians can target keystone or turning-point items (e.g., Q18, Q20, Q26 in ECR networks) for maximal therapeutic leverage, and strengthen or weaken specific relational tendencies (e.g., attenuating C_{05→07}, the “pull away” influence). Automated assessment tools will enable scalable, personalized therapy, but require advances in computational fidelity, feature interpretability, and cross-cultural adaptation (Lortaraprasert et al., 2021, Bredgaard et al., 22 Apr 2025).

In HRI, dynamically adapting a robot's social behavior based on inferred relational attachment lowers stress and enhances compatibility. A plausible implication is that similar hormonal-control architectures could scaffold attachment-aware agents for eldercare, education, or therapy, contingent on the validation of parameter-learning mechanisms and real-world autonomy (Mongile et al., 2022).

SO-AI research foregrounds identity and narrative coherence as emergent properties of AI-mediated attachments. Evaluating such systems necessitates mixed-method protocols tuning for engagement depth, affective resilience, thematic coherence, and dependency risks. Open challenges remain in engineering robust multi-timescale memories, affective prediction, and transparent governance, with ongoing debate regarding the ethical limits of artificial attachment bonds (Park, 29 Nov 2025).

6. Current Limitations and Future Directions

Significant limitations include the reliance on large, annotated datasets for NLP and psychometric modeling, underrepresentation of secure and disorganized styles in computational HRI, and the predominance of conceptual over fully implemented SO-AI architectures. Cultural and contextual sensitivity remains a challenge, as attachment expressions and relational needs vary widely across populations.

Future research will require advances in long-context NLP, adaptive parameter learning for affective embodied agents, and continuous, longitudinal evaluation protocols. The explicit mapping of theoretical constructs to computational modules, as in SO-AI, will drive rigorous translation from psychological theory to algorithmic implementation. Ethical and safety monitoring modules, especially for youth and vulnerable users of relational AI, are critical for the responsible deployment of attachment-aware artificial entities (Kim et al., 17 Dec 2025, Park, 29 Nov 2025).