Graphical-Probabilistic Modeling of Generative Flows in LLM-Native Software Systems

The article introduces a framework called Generation Networks, which utilizes graphical probabilistic models to document generative flows and properties of LLM-native software systems. This approach aims to enhance the rigor of LLM development by enabling principled reasoning and analysis, addressing the stochastic and prompt-dependent behavior of large language models. By formalizing the design and interaction of LLMs, this framework could significantly improve the engineering practices surrounding LLM-native applications, moving beyond heuristic methods.