LLMs Program Tabular Foundation Models Without Gradient Optimization
AI-generated summary A new arXiv paper reports that GPT-6 Astra can program statistical computations in tabular foundation models by manipulating weights and caches without gradient-based optimization. Using zero-initialized weights, Astra achieved a mean ROC AUC of 0.849 across 20 TabArena datasets compared to 0.876 for a pretrained baseline.