Large Language Models as Optimizers: A Survey of Direct vs. Tool-Augmented Approaches and Their Performance Frontiers

In three linesSurvey of LLMs as mathematical optimizers across three paradigms: direct optimization (iterative prompting), tool-augmented optimization (translating to formal specs), and tool-creating optimization (discovering reusable algorithms). Identifies critical reasoning gap and proposes trade-offs between future potential and auditability.

Summary generated by Claude — human-verified