Can LLMs tune search engines better than traditional optimization?

Modern retrieval systems have many interdependent tuning knobs that traditional optimizers (like Bayesian methods) can't handle because they assume independence. Researchers used an LLM agent that learns from full optimization history to navigate these coupled parameters across explore-exploit-refine phases. On HICO-DET benchmarks, it beat standard methods by 33% and improved throughput 15×, with gains strongest when parameters are most entangled. The approach transferred across different vector databases without retraining.