When LLMs Read Tables Carelessly: Measuring and Reducing Data Referencing Errors

Imagine you're building a system that relies on large language models (LLMs) to pull information from tables. You expect these models to accurately reference data, but they often make mistakes, like citing the wrong values or missing important ones altogether. This can lead to incorrect conclusions, especially when the model's reasoning steps are based on faulty data. These mistakes are known as data referencing errors (DREs), and they can undermine the reliability of the model's outputs. Unfortunately, previous research has only scratched the surface of this issue, focusing on small-scale analyses that don't capture the full scope of the problem.

In this paper, the authors take a deeper dive into DREs, systematically evaluating how they occur across various models and tasks. They find that these errors are prevalent in models ranging from 1.7 billion to 20 billion parameters. To tackle this issue, they introduce a new approach that incorporates a critic model specifically designed to detect and filter out these referencing errors. This critic model, which is lightweight at 4 billion parameters, has been shown to improve answer accuracy by up to 12% by effectively identifying both in-distribution and out-of-distribution DREs.

What sets this work apart from prior studies is its comprehensive evaluation and the introduction of a practical solution that enhances the reliability of LLMs when dealing with tabular data. For anyone building systems that depend on accurate data referencing, this approach offers a promising way to improve the overall performance and trustworthiness of their models.

This work systematically evaluates a previously underexplored issue in LLMs, offering new insights into data referencing errors.

The study includes a comprehensive evaluation across multiple models and tasks, supporting its claims with solid experimental results.