New AI Framework Makes Error Detection in Data Transparent and Reliable

Most modern approaches to finding bad data rely on large language models that work as black boxes—they make a judgment call but can't easily explain their reasoning. A new research framework flips this model, using LLMs to construct transparent decision trees that lay out exactly how they evaluate each cell in a spreadsheet.

The Core Innovation#

The researchers developed TreeED, which prompts an LLM to generate a decision tree tailored to the structure and content of a dataset. Rather than simply issuing a yes-or-no verdict, each tree shows its work through a path from root to leaf, with stops along the way to check for errors. The tree nodes fall into three categories: basic rule nodes that validate format and range constraints; Graph Neural Network nodes that detect complex patterns like functional dependencies; and leaf nodes that render the final verdict.

Building on this foundation, they introduced ForestED, which creates multiple decision trees from different row subsets and uses statistical methods to weigh their reliability and reach a consensus judgment. This ensemble strategy mirrors how a panel of judges might work better than a single expert.

Why This Matters#

Data quality is foundational to modern analytics—corrupted or inconsistent values can poison downstream analysis and models. Current methods that treat error detection as a black-box task fail to explain their reasoning and can produce different results on multiple runs, undermining trust and reproducibility.

The new framework addresses these gaps by making the detection process explainable and robust. Organizations can audit how a system catches errors, and the ensemble consensus improves reliability by reducing dependence on any single inference pass.

Results and Performance#

Experiments show the framework achieves a 16.1% average improvement in F1-score over existing best methods, indicating both better precision and recall in catching genuine errors while avoiding false alarms. The approach balances accuracy with interpretability—a notable win in a field often forced to trade one for the other.