CogniPair: GNWT-inspired cognitive architecture for generative agents for Social Pairing - Dating & Hiring Applications

ICLR 2026 Conference Submission14827 Authors

Keywords: Multi-Agent Architecture, Digital Twins, Large Language Models

TL;DR: We operationalize Global Workspace Theory to create the first psychologically authentic AI agents with evolving personalities, demonstrating human-like behavior in large-scale dating simulations with 551 agents based on real-world datasets.

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 strong 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.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 14827