CogniPair: From LLM Chatbots to Conscious AI Agents - GNWT-Based Multi-Agent Digital Twins for Social Pairing - Dating & Hiring Applications
Wanghao Ye 1 , Sihan Chen 2 , Yiting Wang 1 , Shwai He 1 , Bowei Tian 1 , Guoheng Sun 1 , Ziyi Wang 1 , Ziyao Wang 1 , Yexiao He 1 , Zheyu Shen 1 , Meng Liu 1 , Yuning Zhang 1 , Meng Feng 1 , Yang Wang 3 , Siyuan Peng 1 , Yilong Dai 4 , Zhenle Duan 1 , Hanzhang Qin 5,∗ , Ang Li 1,∗ · 2025
Read CogniPair: From LLM Chatbots to Conscious AI Agents - GNWT-Based Multi-Agent Digital Twins for Social Pairing - Dating & Hiring Applications by Wanghao Ye 1 , Sihan Chen 2 , Yiting Wang 1 , Shwai He 1 , Bowei Tian 1 , Guoheng Sun 1 , Ziyi Wang 1 , Ziyao Wang 1 , Yexiao He 1 , Zheyu Shen 1 , Meng Liu 1 , Yuning Zhang 1 , Meng Feng 1 , Yang Wang 3 , Siyuan Peng 1 , Yilong Dai 4 , Zhenle Duan 1 , Hanzhang Qin 5,∗ , Ang Li 1,∗ (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-03543.txt.
Introduction
Wanghao Ye and a large team present CogniPair, a platform of digital twins meant to rehearse social encounters before they happen. Ordinary chatbots lack the inner workings of human psychology, so the team built agents on Global Workspace Theory, a model of consciousness. Specialized sub-agents handle emotion, memory, social norms, planning, and goals, coordinated through a shared workspace.
A faithful twin also needs the right personality, so the team designed an adventure-style test that reads character from choices in story scenes rather than from self-descriptions. Twins then meet in simulated dates and job interviews, letting both sides gauge fit. Tests with 551 agents and a university speed-dating dataset showed a 72 percent correlation with human attraction patterns and about 78 percent accuracy at predicting matches.
Catalog record
| CREATOR | Wanghao Ye 1 , Sihan Chen 2 , Yiting Wang 1 , Shwai He 1 , Bowei Tian 1 , Guoheng Sun 1 , Ziyi Wang 1 , Ziyao Wang 1 , Yexiao He 1 , Zheyu Shen 1 , Meng Liu 1 , Yuning Zhang 1 , Meng Feng 1 , Yang Wang 3 , Siyuan Peng 1 , Yilong Dai 4 , Zhenle Duan 1 , Hanzhang Qin 5,∗ , Ang Li 1,∗ |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 117 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-03543
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CogniPair: From LLM Chatbots to Conscious AI Agents - GNWT-Based Multi-Agent Digital Twins for Social Pairing - Dating & Hiring Applications Wanghao Ye 1 , Sihan Chen 2 , Yiting Wang 1 , Shwai He 1 , Bowei Tian 1 , Guoheng Sun 1 , Ziyi Wang 1 , Ziyao Wang 1 , Yexiao He 1 , Zheyu Shen 1 , Meng Liu 1 , Yuning Zhang 1 , Meng Feng 1 , Yang Wang 3 , Siyuan Peng 1 , Yilong Dai 4 , Zhenle Duan 1 , Hanzhang Qin 5,∗ , Ang Li 1,∗ Abstract Current large language model agents lack authentic human psychological processes necessary for genuine digital twins. We present the first computational implementation of Global Workspace Theory (GNWT), creating agents with multiple specialized sub-agents (emotion, memory, social norms, planning, goal-tracking) coordinated through a global workspace broadcast mechanism. This architecture allows agents to maintain consistent personalities while evolving through social interaction. Our CogniPair simulation platform deploys 551 GNWT-Agents for speed dating interactions, grounded in real data from the Columbia University Speed Dating dataset. Evaluations show unprecedented psychological realism, with agents achieving 72% correlation with human attraction patterns and outperforming baselines in partner preference evolution (72.5% vs. 61.3%). Human validation studies confirm our approach’s fidelity, with participants rating their digital twins’ behavioral accuracy at 5.6/7.0 and agreeing with their choices 74% of the time. This work establishes new benchmarks for psychological authenticity in AI systems and provides a foundation for developing truly human-like digital agents. 1 University of Maryland, College Park 2 Viterbi School of Engineering, University of Southern California 3 Northeastern University 4 Department of Computer Science, University of Florida 5 National University of Singapore ∗ Joint corresponding authors University of Maryland Departments: Electrical and Computer Engineering, Computer Science, Robert H. Smith School of Business, School of Public Policy, Department of Agricultural and Resource Economics {wy891, yiting, bowei, guoheng, ziyi, ziyao, yexiao, zheyu, meng, yuning, mfeng, siyuan, zhenle, angli}@umd.edu [email protected] [email protected] [email protected] [email protected] 1 Introduction Human social interactions—from dating to job interviews—require not just coherent dialogue but authentic psychological processes including emotion regulation, memory consolidation, and dynamic preference formation. LLM-based agents have been applied to model human social interactions, showing promise in domains such as customer service, healthcare assistance, and educational tutoring Park2023 ; Xu2023 ; Wang2024 . Despite recent advances, current LLM-based agents face two fundamental limitations that restrict their ability to model human behavior realistically: (1) the psychological behavior gap —they cannot authentically simulate internal mental states, emotional processing, or evolving preferences Zhang2024b ; Serapio-Garcia2023 ; Jiang2023 ; Guo2024 ; and (2) the social behavior gap —they fail to capture the complex dynamics of human-to-human interactions where preferences and behaviors co-evolve through social experiences Park2023 ; Zhang2018 ; Huang2024 ; Park2022social ; Aher2023 ; Sun2024 . The psychological behavior gap manifests in two critical problems: the individualization problem , where agents act like generic humans rather than specific individuals with unique psychological profiles, and the static personality problem , where agents cannot evolve mentally through experience. Existing approaches such as Stanford’s Generative Agents Park2023 demonstrated emergent behaviors but relied on fictional personas without real human data. PersonaChat Zhang2018 introduced personality descriptions that remain synthetic and fixed. Recent personality modeling efforts Serapio-Garcia2023 ; Jiang2023 achieve only surface-level behavioral mimicry without cognitive grounding. Most critically, these approaches treat personality as immutable prompts rather than dynamic psychological states shaped by experience. The social behavior gap emerges when attempting to model authentic social interactions, particularly in complex domains such as relationship formation. Current LLM agents lack the capability to engage in authentic social dynamics where preferences evolve through interaction, emotional responses adapt to social feedback, and behavioral patterns shift based on interpersonal experiences. This limitation becomes particularly pronounced in domains requiring complex social cognition, such as dating scenarios where mutual attraction emerges through dynamic, bidirectional assessment processes. Figure 1 : Comparison of Cognitive Architectures: Traditional RL/Chatbot (left), GNWT Agent Architecture (center), and GNWT Human Model (right), showing the evolution from linear processing to brain-inspired parallel processing with global workspace integration. To address these fundamental gaps, we turn to Global Workspace Theory (GNWT) Mashour2020 ; Baars2005 , a leading neurocognitive model that explains how human consciousness emerges from the interaction of specialized brain modules. In human cognition, GNWT describes how disparate neural processes—emotion, memory, perception, and planning—compete for access to a central "global workspace." When information becomes sufficiently salient, it triggers a broadcast that propagates this content throughout the brain, creating our unified stream of consciousness. This theoretical foundation provides a clear roadmap for building agents that can overcome both the psychological and social behavior gaps(Figure 1 ). To address the psychological behavior gap , we operationalize GNWT into a computational agent architecture where each individual agent contains multiple specialized sub-agents working in parallel as a unified consciousness. Our GNWT-Agent implements five specialized cognitive modules—Emotion, Memory, Planning, SocialNorms, and GoalTracking—each grounded in neurocognitive theories and parameterized by the agent’s Five-Factor personality profile. By implementing GNWT’s broadcast mechanism computationally, we create agents with genuine internal psychological dynamics—emotion sub-agents generate affective responses, memory sub-agents consolidate experiences, social norms sub-agents manage cultural awareness, planning sub-agents develop strategies, and goal-tracking sub-agents maintain objectives. This architecture fundamentally differs from traditional LLM agents that process inputs sequentially without internal state evolution. To bridge the social behavior gap , we developed CogniPair , a social-influence decision system that enables GNWT-Agents to engage in authentic social interactions and evolve through experience. CogniPair is not merely a testbed but a comprehensive system for modeling and guiding social influence between individuals, ultimately optimizing decision processes across various social contexts. While our primary evaluation uses a speed dating testbed, the CogniPair system itself can be extended to other social decision environments such as team formation, negotiation scenarios, and collaborative problem-solving. We selected speed dating as our evaluation domain because it exemplifies the most challenging aspects of human social cognition—rapid compatibility assessment, dynamic preference formation, emotional regulation under uncertainty, and integration of multiple information streams. The Columbia University Speed Dating dataset Fisman2006 ; Fisman2008 provides rich behavioral ground truth, including pre- and post-interaction preferences, attraction ratings, and decision outcomes, enabling rigorous evaluation of social realism. Our dating testbed deploys 551 GNWT-Agents in a two-level simulation architecture: the internal level models psychological processes within each agent, while the external level simulates social dynamics between agents. This scale—20 times larger than previous personality-based simulations Park2023 —enables statistically valid analysis of emergent social patterns, creates sufficient diversity for complex relationship networks, and allows measurement of population-level phenomena. Through this dual approach—addressing the psychological gap with GNWT-Agents and the social gap with the CogniPair system—our framework uniquely enables both individualization and dynamic evolution. The global workspace mechanism naturally handles the stability-plasticity dilemma in personality modeling: core traits remain stable through persistent attention patterns while allowing adaptive changes through experience-driven broadcast priority shifts. Unlike previous approaches that treat personality as static prompts or surface behaviors, our architecture models the cognitive processes underlying personality and the social dynamics shaping its evolution. Our evaluation framework measures psychological and social realism across multiple dimensions. For psychological realism, we assess preference consistency (how well agents maintain core values while adapting), emotional coherence (whether affective responses follow human psychological principles), and