Liping Huang

We propose a training-free hybrid framework where LLMs generate high-level goals via prompt-based harmony search and optimizers enforce constraints in dynamic ride-hailing, outperforming RL, manual decomposition, and LLM-only baselines.