Why Large Language Models Fail at Tabular Prediction

(arxiv.org)

23 points | by sbulaev 3 hours ago

4 comments

  • _joel 2 minutes ago
    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.
  • ismailmaj 53 minutes ago
    Unsure if it's LLMs that fail at tabular data or its just that tree boosting are spectacular at that task.
  • heaney-555 15 minutes ago
    >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.

  • cyanregiment 1 hour ago
    Just have 2 LLMs debate whether tabs or spaces are the superior choice