ProgramTab boosts LLM table reasoning via Python preprocessing

The authors propose ProgramTab, a framework that guides large language models to perform tabular data preprocessing using Python code for in-context learning.

• The approach addresses performance degradation in large tables caused by long text modeling difficulties and input length limits.
• It overcomes the structural inconsistencies of web tables that make SQL queries unsuitable for mathematical logic operations.
• ProgramTab combines row and column extraction with SQL generation to extract momentous contents from tabular data.
• Experiments on table reasoning datasets show it outperforms all LLM-based baselines.

The framework effectively handles table-based reasoning tasks by leveraging programmatic paradigms instead of relying solely on direct text-to-SQL conversion.