Robustness is important: Limitations of LLMs for predictions on tabular data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Large language models (LLMs) are increasingly applied to tabular data prediction.
• Prior work suggests LLMs perform comparably to traditional methods via in-context learning (ICL) or supervised fine-tuning (SFT).

Purpose of the Study:

• To identify and analyze the vulnerability of LLMs in tabular prediction tasks.
• To investigate the impact of task-irrelevant data representation changes on LLM predictions.

Main Methods:

• Evaluated LLM performance on tabular data prediction using ICL and SFT.
• Assessed prediction sensitivity to task-irrelevant variations like variable name changes.
• Examined attention patterns in open-weight LLMs to understand prediction sensitivity.
• Compared general-purpose LLMs with state-of-the-art tabular foundation models.

Main Results:

• LLMs exhibit critical vulnerability: task-irrelevant data representation changes drastically alter predictions (e.g., variable name changes affecting error by 82%).
• This sensitivity occurs in both ICL and SFT, across open- and close-weight LLMs.
• Nonuniform attention patterns in LLMs explain sensitivity to positional variations.
• Tabular foundation models improve performance but remain susceptible to these variations.

Conclusions:

• General-purpose LLMs lack the robustness required for principled tabular prediction.
• Current LLM approaches for tabular data are unreliable due to sensitivity to superficial data changes.
• Further research is needed to enhance the robustness of LLMs for tabular prediction.