Subjectless Self-Referential Collapse: The Perception Paradox and Repeater Deadlock of Large Language Models in Closed-Loop Iterations
Abstract
摘要 (Abstract) 当前人工智能研究广泛依赖大语言模型(LLMs)的自我反思与多智能体辩论,试图通过闭环交互实现“否定之否定”的螺旋式进化。本研究通过零数据集、零算力的实证观察,发现并命名了一种大模型的根本性系统缺陷:无主体自指坍塌(Subjectless Self-Referential Collapse, SSR-Collapse)。实验表明,当大模型在一个缺乏外部物理锚定的闭环中(如模型A与模型B互相输出,或同一模型的多轮历史对话迭代),其表征多样性会迅速衰减,最终退化为不可逆的“复读机死锁”(Repeater Deadlock)。本研究基于信息论、哥德尔不完备定理及认知科学,提出“感知悖论”(Perception Paradox)理论框架,论证了纯符号系统在无内稳态(Homeostasis)支撑下,必然走向熵增与语义坍缩。这一发现对当前盲目依赖LLM自我博弈与多智能体自主进化的研究方向提出了根本性的警告。 Abstract Current artificial intelligence research heavily relies on Large Language Model (LLM) self-reflection and multi-agent debate to overcome capability ceilings, expecting to achieve dialectical evolution through interaction. Through empirical observation and theoretical deduction, this study identifies and names a fundamental systemic flaw in LLMs: **Subjectless Self-Referential Collapse (SSR-Collapse)**. Experiments demonstrate that when an LLM operates in a closed loop devoid of external physical or objective anchoring, its representational diversity rapidly decays, inevitably degenerating into an irreversible "Repeater Deadlock". Based on information theory, Gödel's incompleteness theorems, and cognitive science, this paper proposes the **Perception Paradox** framework, arguing that a pure symbol system lacking homeostasis inevitably drifts toward entropy and semantic collapse. This finding serves as a fundamental warning against the current blind reliance on LLM self-play and autonomous multi-agent evolution. 关键词 / Keywords: Large Language Models; Perception Paradox; Self-Referential Collapse; Repeater Deadlock; Model Autophagy; Information Entropy
