Tag
This paper systematically evaluates data referencing errors (DREs) in LLMs processing tables and proposes critic-based filtering and rejection sampling to improve answer accuracy. A lightweight 4B-parameter critic model achieves high detection accuracy for DREs.