Emergence of High-Dimensional Communication Protocols in Multi-Agent Systems: A Non-Anthropocentric Analysis
This is a paper based on my thesis.
Abstract: The emergence of communication protocols between artificial agents has been extensively studied, yet traditional approaches often constrain these protocols to human-interpretable patterns or reduced dimensionality spaces. We present a novel investigation into how artificial agents naturally develop communication protocols in high-dimensional spaces, challenging traditional anthropocentric constraints. Our methodology combines multiple analytical approaches including information theory, manifold learning techniques, and geometric analysis to study communication patterns in a 768-dimensional embedding space. We employ two language models (GPT-2 and DistilGPT-2) in a multi-agent reinforcement learning environment, analyzing their emergent communication through sophisticated dimensionality estimation techniques and statistical validation methods. Key findings reveal three fundamental properties: (1) agents develop a highly efficient communication protocol utilizing only 5.5-6.5 dimensions of the available space, achieving a stable channel capacity of 2.71 bits; (2) communication develops through distinct stages: exploration (0-10000 steps), consolidation (10000-30000 steps), and maturation (30000+ steps); and (3) the protocol exhibits sophisticated geometric properties including positive Ricci curvature (mean 0.514) and fractal dimension (2.646), indicating hierarchical organization. These results demonstrate that artificial agents, when unconstrained by human-centric communication requirements, develop efficient, structured protocols optimized for their architectural capabilities. Our findings have significant implications for multi-agent system design, inter-model communication, and our understanding of how intelligence might naturally emerge in artificial systems.