效率衰减现象：对思想语言假说的计算挑战

This study presents a computational model investigating whether thought necessitates a language-like format, as proposed by the Language of Thought (LoT) hypothesis. The 'AI Private Language' thought experiment examines whether two artificial agents can develop an efficient and inscrutable communication protocol through multi-agent reinforcement learning (MARL). When forced to use a human-comprehensible language, their performance declines, leading to the Efficiency Attenuation Phenomenon (EAP), which challenges the LoT. In a cooperative navigation task under partial observability, results indicate that agents utilizing an emergent protocol achieve 50.5% higher efficiency than those using a predefined human-like symbolic protocol, confirming the existence of EAP. This finding suggests that optimal communication protocols may not require the structure of human language.