An important Bottleneck in LLM Optimizers: Principle-Level Local Optima

Nian Ran, Tianhe Jia, Zhengyang Mu +6

Feedback-driven LLM optimizers are increasingly used for evolutionary search, Bayesian optimization, heuristic design, and scientific discovery. Yet rapid initial gains often give…