Large language models as knowledge-guided proposers for reliability-based design optimization of truss structures

Large language models (LLMs) have shown remarkable performance across a wide range of open-domain tasks, and a small but growing body of recent work has begun to employ them in structural optimization. This study extends that work to truss structures, examining how an LLM behaves when it is embedded in the reliability-based design optimization (RBDO) loop of a truss. We propose a closed-loop framework in which the language model acts as a knowledge-guided proposer of candidate designs. Through carefully designed prompts and in-context learning, the framework reads the structural information and the accumulated search history and proposes cross sections, which are accepted only if they satisfy both the deterministic and the probabilistic constraints under a finite element solve and a first-order reliability evaluation. The approach is assessed on four benchmarks, the 10-bar, 25-bar, and 120-bar trusses and the real 137-bar Burro Creek bridge, against two metaheuristics, one deterministic, and one surrogate-assisted optimizer under identical settings. The framework reaches feasible designs of weight broadly competitive with the metaheuristic baselines while requiring about 14 to 47 times fewer coupled structural and reliability evaluations than the metaheuristics at their native budgets, although at an equal evaluation budget it is lighter than the best classical baseline only on the dome, and its contribution is clearest on structures whose material must be apportioned among several groups of distinct structural role. Furthermore, an investigation of how structural knowledge is written into the prompt, of the surface wording, and of three different language models shows that the choice of model is among the largest sources of variation, while the framework is less repeatable than the classical optimizers. By framing the language model as a single component within a loop whose feasibility and reliability checks rest on deterministic mechanics rather than on the model itself, this study provides evidence that such a model can act as a competent proposer within a verified loop.