MALMAS: Memory-Augmented Multi-Agent System for Automated Feature Generation

A recent paper published on arXiv introduces MALMAS, a Multi-Agent System that utilizes Memory-Augmented LLMs for the automated generation of features from tabular data. This system overcomes the shortcomings of conventional techniques that depend on static operator libraries and fail to utilize task semantics. Although recent LLM-based methods have emerged, they are still hindered by a limited feature space due to their fixed generation patterns and inadequate feedback from learning objectives. MALMAS enhances the generation process by assigning specific roles to agents, with a Router Agent selecting an appropriate subset for each iteration to expand exploration. Additionally, memory augmentation is incorporated to enhance feature quality. The research can be accessed at arXiv:2604.20261.