Embodied Consciousness-Like Emergence in a Multi-Agent Enclosed Arena v1.0 (Draft)
Xinping X. Wang
Zenodo (CERN European Organization for Nuclear Research) April 16, 2026 preprint DOI: 10.5281/zenodo.19605082 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Proposes an embodied multi-agent framework using the MuJoCo physics engine to examine whether LLM-driven agents display alliance, deception, betrayal, altruism, self-referential expression, and moral collapse under survival pressure across four escalating phases. The authors state they do not claim phenomenal consciousness in AI, instead exploring operationally how embodiment, scarcity, and social interaction shape self-modeling and moral adaptation. |
Abstract
This paper proposes an embodied multi-agent experimental framework based on the MuJoCo physics engine to investigate whether LLM-driven agents can exhibit complex social behavior and consciousness-like emergent behavior under survival pressure. The system embeds LLMs into a closed arena where agents are given physical bodies, finite energy, differentiated personalities, asymmetric resources, and incomplete information. Across four progressively intensified phases—peace, resource contraction, famine crisis, and final judgment—the framework examines the emergence of alliance, deception, betrayal, altruism, self-referential expression, and moral collapse. The paper does not claim phenomenal consciousness in AI; instead, it operationally explores how embodiment, scarcity, and social interaction shape self-modeling, moral adaptation, and spontaneous strategic behavior in LLM-based agents.