Researchers from IBM and ETH Zurich have developed an AI model that can directly extract key-value pairs from document images without the need for optical character recognition (OCR) preprocessing. The model, SmolDocling, is a compact vision-language model that performs identification, localization, and association in a single pass. To address data limitations, the researchers designed an augmentation pipeline combining synthetic form filling and graph-based crops. They also introduced a layout-aware evaluation framework to assess the model's performance. On several datasets, including FUNSD and XFUND, SmolDocling outperformed larger zero-shot VLM baselines while being significantly smaller in size. --- Why it matters: This matters because it can improve document processing pipelines by reducing error propagation and increasing efficiency. It also has potential applications in areas such as data entry, document management, and information retrieval. Source: https://arxiv.org/abs/2608.20868