Reproduce “TabClean: Reusable LLM-Synthesized Programs for Tabular Data Cleaning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yibo Wang, Riteng Zhang, Yinghao He, Yongye Su, Bharat Bhargava, Chunwei Liu

TABCLEAN proposes a system that synthesizes reusable, guarded cleaning programs using Large Language Models (LLMs) from small annotated development sets. This approach significantly improves cleaning effectiveness, achieving F1 scores up to 0.99, while also demonstrating orders of magnitude faster runtime and reduced API costs compared to existing LLM-based methods.

Key insights

  • Decomposing the cleaning task into a finite-state machine with role-specialized LLM agents and selective, role-aware memory is crucial for effective LLM-guided program synthesis, improving quality and efficiency.
  • Tracking the 'best validated program' is essential during iterative LLM synthesis to prevent performance regressions, as LLM-generated programs do not always improve monotonically.
  • Synthesizing guarded, deterministic Python programs amortizes LLM inference costs and provides scalable, auditable solutions that significantly outperform direct LLM inference in terms of cost and runtime.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.