ARASH: Adaptive Retrieval and Shot Selection for Tabular Prediction

A recent study published on arXiv (2608.17856v1) presents ARASH (Adaptive, query-specific Retrieval And Shot selection), a technique designed to enhance the efficiency of Tabular Foundation Models (TFMs) such as TabPFN. Traditionally, TFMs demand substantial computational power and regular retraining. ARASH utilizes in-context learning via few-shot prompting, identifying the best shots through an analysis of local neighborhoods in the training dataset. This approach significantly decreases prompt length and memory consumption by a factor of 1261.5, providing a more resource-efficient solution for tasks involving tabular predictions.