Tag
This paper introduces a comprehensive benchmark for evaluating LLMs in key-value extraction from documents under OCR noise, revealing substantial performance degradation and emphasizing the need for joint optimization of OCR quality and LLM reasoning.