LATTEArena: An Evaluation Framework for LLM-powered Tabular Feature Engineering (Extended Version)

LATTEArena is a newly introduced benchmarking framework designed for LLM-powered Automated Tabular Feature Engineering (LATTE), addressing the lack of standardized evaluation tools in the field. It deconstructs 15 LATTE methods into a 6-dimensional taxonomy and evaluates 24 configurations across 7 research questions, focusing on metrics beyond predictive accuracy, including token efficiency and execution robustness. This framework facilitates systematic comparisons and provides resources for optimal deployment, with all code and datasets made publicly accessible, thus promoting community engagement in the benchmarking process.