CogniPair: Neurocognitive AI for Social Pairing

• CogniPair Platform is a neurocognitive multi-agent AI system that operationalizes Global Workspace Theory to generate authentic digital twins for social pairing applications.
• It integrates specialized cognitive modules such as emotion, memory, planning, social norms, and goal-tracking, coordinated by a global workspace for dynamic agent interactions.
• The platform employs adventure-based personality assessments and rigorous empirical validation to replicate human social dynamics in dating and hiring scenarios.

The CogniPair Platform is a neurocognitive multi-agent AI system that operationalizes Global Workspace Theory (GNWT) within LLM agents, enabling the creation of psychologically authentic digital twins for social pairing applications, notably in the domains of dating and hiring. This platform addresses key limitations of traditional LLM chatbots by integrating specialized cognitive modules—emotion, memory, planning, social norms, and goal-tracking—coordinated by a computational global workspace. In conjunction with an adventure-based personality assessment for robust initialization, CogniPair simulates bidirectional human-agent social interactions and provides quantifiable validation against human datasets. The following sections detail its architecture, assessment methodology, social application framework, validation approach, and implications for psychological realism in artificial agents.

1. Neurocognitive Agent Architecture: GNWT-Based LLM Agents

CogniPair implements the Global Neuronal Workspace Theory by structuring each agent (“GNWT-Agent”) as a composite of five specialized cognitive modules: Emotion, Memory, Planning, SocialNorms, and GoalTracking. These modules operate in parallel, each providing a functional analogue of specific neural systems implicated in human cognition (e.g., limbic system, hippocampus, DLPFC, mPFC, OFC). Their coordination is mediated by a global workspace integration mechanism, which enforces salience-driven broadcasting, conflict detection/resolution, and the construction of the agent’s final output.

Mathematically, a GNWT-Agent is defined as

$$\mathcal{E} = (\mathcal{M}, \mathcal{W}, \mathcal{I}, \mathcal{L}, \mathcal{P})$$

where $\mathcal{M}$ is the set of modules (each parameterized by the agent’s Big Five trait vector), $\mathcal{W}$ the global workspace, $\mathcal{I}$ an information bottleneck, $\mathcal{L}$ the LLM interface, and $\mathcal{P}$ the personality profile space. For each input, modules return candidate outputs with associated salience scores $s_i$, and the global workspace processes:

• Parallel module response computation:

$$R_{\text{Module}} = f_{\text{Module}}(Q, H, \mathrm{GW}, \text{Trait Parameter})$$

• Aggregation via weighted integration:

$$\text{Response} = \sum_{M} \alpha_M(P) \cdot R_M + \beta(P) \cdot G(\mathrm{GW})$$

where $\alpha_M(P)$ weights each module’s contribution by the agent’s trait vector.

This modular neuro-symbolic design grants each agent a persistent, evolvable internal state enabling emotion regulation, episodic and semantic memory, explicit social norm adherence, forward planning, and dynamic goal management—capabilities that are absent in standard single-prompt LLM agents.

2. Adventure-Based Personality Initialization

Authentic digital twin construction in CogniPair depends upon high-fidelity personality initialization, which is achieved via a novel adventure-based personality assessment. This test replaces traditional self-report questionnaires with 12–15 interactive scenarios (“adventures”) that elicit behavioral choices and open-ended responses in contextually rich, simulated settings.

Key design features include:

• Scenario adaptation for trait confidence: The system adaptively selects subsequent scenarios to refine weak or conflicting evidence regarding particular Big Five dimensions (openness, conscientiousness, extraversion, agreeableness, neuroticism).
• Analysis via LLM-based inference: Each user response is analyzed using LLaMA 3 locally or GPT-4o in the cloud, continually updating trait estimates and confidence levels until a stopping criterion is reached.
• Social desirability bias avoidance: By inferring personality from hypothetical choices rather than explicit self-ascription, this method reduces self-presentation effects (validated by $r=0.82$ correlation with Big Five instruments, Cohen’s $d=0.74$ lower social desirability bias in empirical studies).
• Open-ended behavioral sampling: Scenarios target diverse real-world behaviors (e.g., conflict resolution, leadership, ethical dilemmas, social initiative) to produce robust, context-dependent digital twin initialization.

The resulting trait profile forms the basis for each agent’s internal parameterization and cognitive dynamics.

3. Simulated Social Pairing Framework

CogniPair deploys GNWT-Agents—each initialized as a digital twin of a real or hypothetical user—in multi-turn, realistic simulations of social interactions. Two primary platforms are supported:

• Romantic Pairing: Agents engage in simulated speed-dating sessions mapped to protocols from the Columbia University Speed Dating dataset (8 turns, 4 minutes per “date”). Each agent tracks not only conversational content, but also evolving preference vectors, memory traces, emotional tone, and social norm adherence.
• Hiring/Job Interviews: Candidate and employer agents participate in simulated interview scenarios, encompassing technical, behavioral, and cultural fit dimensions.

Importantly, CogniPair enforces bidirectional cultural fit assessment: both parties must independently decide to accept or reject, mirroring mutual selection in human social settings. The platform records module-level activations and internal rationales, providing interpretability into the features underlying compatibility decisions.

4. Empirical Validation and Evaluation

Supervised validation leverages re-creation of all 551 participants from the Columbia University Speed Dating dataset as GNWT-Agents, with simulation protocols precisely mirroring real-world conditions (including conversational context, turn order, and duration). Comparison baselines include prompt-based LLMs, memory-augmented agents, and debate-based multi-agent systems.

Principal findings include:

• Partner Preference Evolution: GNWT-Agents replicate the empirical shift of human preference patterns over the course of interaction (e.g., attractiveness +25.0% for agents vs. +39.0% for humans; intelligence −15.2% for agents vs. −24.8% for humans), with correlation coefficients ranging $0.73$–$0.86$.
• Self-Perception Adaptation: Post-interaction self-perception adjustment aligns between agents and humans ($r=0.81$–$0.82$).
• Match Prediction: GNWT-Agents achieve $77.8\%$ match prediction accuracy (baseline: $69.1\%$; static traits only: $58.9\%$).
• Behavioral Authenticity: Human-agent behavioral agreement is measured at $74\%$ for speed-dating and $81\%$ for job interviews, with trait correlation approximately $0.83$ and conversational authenticity ratings of $5.4$–$5.6/7.0$.

These results demonstrate that CogniPair not only achieves quantitative alignment with human social dynamics, but also exhibits emergent phenomena observed in real populations.

Summary Table: Core Innovation Results

Metric	Speed Dating	Job Interview
Behavioral fidelity	5.6/7.0	5.8/7.0
Decision concordance (human-agent)	74%	81%
Personality trait correlation	0.83	0.81
Match prediction accuracy	77.8%	—

5. Implications for AI Psychological Authenticity

CogniPair demonstrates that integrating a global workspace architecture and explicit neuro-symbolic modules into LLM agents enables the emergence of distinct, stable, and behaviorally dynamic digital twins. The platform models not only static trait differences but also in-interaction adaptation, social feedback effects, and bidirectional cultural fit—phenomena previously unattainable with classic LLM-based conversational AI.

By surpassing the “generic/unchanging agent” limitation and enabling platform-level behavioral realism (including authentic preference evolution and group-level distributional shifts), CogniPair advances the fidelity and interpretability of digital twins for a range of social simulation and decision-support applications.

6. Future Directions and Applications

The CogniPair platform is positioned for extension across various domains:

• Nonverbal and Multimodal Interaction: Incorporating physiological, audiovisual, and paralinguistic signals to extend beyond text-based simulation.
• Cross-Cultural Cognitive Modeling: Refinement of the SocialNorms module for diverse cultural contexts and international deployment.
• Enterprise HR Technology: Scalable, bidirectional, and bias-resistant hiring solutions with team-level fit analysis.
• Network-Level Social Simulation: Modeling emergent behaviors in multi-agent societies, including group formation, negotiation, and collective decision-making.
• Personalized Coaching and Education: Leveraging digital twins for tailored mentorship, counseling, or group facilitation.
• Ethical and Privacy Considerations: Ongoing research into privacy preservation and bias mitigation in digital twin construction and deployment.

A plausible implication is the applicability of the CogniPair framework for longitudinal relationship counseling, remote collaborative team formation, and transparent negotiation or conflict resolution in virtual environments.

---

CogniPair sets a new standard for the construction of psychologically authentic, adaptive, and explainable AI agents, offering a rigorously validated framework for social pairing tasks and providing a foundation for further research in computational social cognition, digital twin development, and human-AI interaction at scale.