From Pixels to Pairs
An academic paper published on arXiv that systematically benchmarks the performance of large language models on document key-value pair information extraction under OCR-noise conditions, and proposes a unified benchmarking protocol for evaluating such tasks.
Timeline (last 90 days)
The paper 'From Pixels to Pairs' benchmarks LLM performance on document information extraction under OCR-noisy conditions.
The paper attributes extraction performance to two dominant factors: semantic reasoning capability over text and the preservation of text fidelity under OCR noise.
The paper recommends joint optimization that simultaneously improves OCR quality, structured reasoning capability, and LLM semantic modeling.
All Facts (3)
The paper 'From Pixels to Pairs' benchmarks LLM performance on document information extraction under OCR-noisy conditions.
50%UnverifiedThe paper attributes extraction performance to two dominant factors: semantic reasoning capability over text and the preservation of text fidelity under OCR noise.
50%UnverifiedThe paper recommends joint optimization that simultaneously improves OCR quality, structured reasoning capability, and LLM semantic modeling.
50%