📄 中文摘要
研究表明，经过自组织临界性预训练的PLDR-LLMs在推理时展现出推理能力。PLDR-LLM在临界状态下的演绎输出特征与二阶相变相似。在临界状态下，相关长度发散，演绎输出达到亚稳态。稳态行为表明，演绎输出从训练数据集中学习到的表示等同于缩放函数、普适性类和重整化群，从而在过程中实现了泛化和推理能力。可以根据模型演绎输出参数的全局统计定义一个序参量，以此来量化PLDR-LLM的推理能力。该模型的推理能力表现出更好的特性。
PLDR-LLMs在自组织临界性下的推理能力
出处: PLDR-LLMs Reason At Self-Organized Criticality
发布: 2026年3月26日
📄 English Summary
PLDR-LLMs Reason At Self-Organized Criticality
The study demonstrates that PLDR-LLMs pretrained at self-organized criticality exhibit reasoning capabilities during inference. The characteristics of the deductive outputs of PLDR-LLMs at criticality resemble those of second-order phase transitions. At criticality, the correlation length diverges, and the deductive outputs reach a metastable steady state. This steady-state behavior indicates that the deductive outputs learn representations equivalent to scaling functions, universality classes, and renormalization groups from the training dataset, leading to enhanced generalization and reasoning capabilities. An order parameter can be defined based on the global statistics of the model's deductive output parameters at inference, quantifying the reasoning capabilities of PLDR-LLMs.