PPDL: LLM-Based Flows as Probabilistic Programs

The paper proposes a probabilistic language that treats LLM-based flows as probabilistic programs, enabling developers to model uncertainty across multiple LLM calls and tool interactions. It aims to provide clear confidence measures and improve accuracy in LLM applications, which currently often lack quantifiable reliability. The approach is designed to help both developers and end-users trust results from complex LLM pipelines. The paper is available on arXiv as 2608.05234v1, indicating ongoing research in this area. This could lead to more robust LLM orchestration frameworks and better uncertainty quantification in production AI systems.