The first thing I'd do if working with an LLM on tabular data is to ask what the best tool would be to work with that data and build up a proper harness to work with the data sensibly. Rawdogging LLM isn't the tool for forecasting like this, as they found.
>We study a frontier LLM in its purest inference regime - a single generation pass over a prompt containing the full training and test data, with no tools, no agentic scaffolding, and no fine-tuning
Sigh. So this is somewhat interesting niche academic research but utterly irrelevant to real-world use cases.
Sigh. So this is somewhat interesting niche academic research but utterly irrelevant to real-world use cases.