TAFFY: A Task-Adaptive Tabular Foundation Model with In-Context Diversity
TAFFY is a new tabular foundation model designed to improve in-context learning by better inferring task-specific relationships from context. It uses an In-Context Diversity Prior to create varied training contexts and a Task-Conditioned Looped Transformer to iteratively refine these contexts based on the specific task, leading to state-of-the-art performance across multiple benchmarks.