MetaLLMiX is a zero-shot hyperparameter optimization framework that combines meta-learning, explainable AI, and efficient LLM inference. To address the trial-and-error and high-cost API challenges of existing AutoML and LLM-based approaches, it leverages SHAP explanations to recommend optimal hyperparameters and pre-trained models based on past experimental results without additional trials. LLMs are used as judges to control output format, accuracy, and completeness. In experiments using eight medical image datasets and nine open-source lightweight LLMs, it achieves competitive or superior performance to existing HPO methods while significantly reducing computational costs. It outperforms existing API-based approaches, achieving optimal results on five out of eight tasks, reducing response times by 99.6-99.9%, and achieving the fastest training times (2.4-15.7x) on six datasets. Accuracy remains within 1-5% of the best-performing benchmark model.
