A paper applies a least action principle to estimate the generative directionality of LLMs inside agents. Measured transition probabilities between generated states are consistent with a tendency toward detailed balance, suggesting LLMs implicitly learn underlying potential functions that may transcend architectures and prompt templates.

— via arXiv — Physics, Zhuo-Yang Song, Qing-Hong Cao, Ming-xing Luo, Hua Xing Zhu