Can large language models find the design point? Benchmarking LLM-based MPP search in structural reliability
Über dieses Buch
Finding the Most Probable Point (MPP), or design point, is a critical step in gradient-based structural reliability methods. Classical solvers such as HL-RF and iHL-RF locate the MPP reliably but often require a large number of limit state function (LSF) evaluations, which becomes prohibitive when the LSF involves expensive computations such as finite element analysis. This paper investigates whether a large language model (LLM) can serve as an autonomous algorithm designer for MPP search, generating problem-adapted search routines from a structured natural language prompt without any task-specific training. The proposed framework, LLMMPP, queries the LLM once per problem and executes the returned algorithm directly on the LSF. The framework leverages DeepSeek-V3, a state-of-the-art open-weight LLM, to generate problem-adapted search routines from a structured natural language prompt without any task-specific training. Experiments on four benchmark problems from the structural reliability literature show that LLMMPP reduces the number of LSF evaluations by a factor of 2.1 to 28.9 relative to the best classical method, while maintaining comparable or superior accuracy in the identified design point. These results indicate that LLMs can be used as efficient, zero-shot MPP optimizers, particularly for nonlinear, multi-modal, and high-dimensional performance functions.
