Unsure if it's LLMs that fail at tabular data or its just that tree boosting are spectacular at that task.