Brief

Researchers have developed MinerU-Popo, a novel framework designed to enhance structured document parsing by addressing limitations in current VLM-based OCR models. This system focuses on reconstructing document-level logical structures, such as paragraphs and tables, that are often fragmented across page boundaries. By employing a lightweight post-processing model fine-tuned on a custom dataset and utilizing dynamic chunking for long documents, MinerU-Popo significantly improves accuracy in RAG applications and reduces latency. AI

Apple researchers have introduced VSAS-Bench, a new framework designed to evaluate visual streaming assistant models in real-time. Unlike previous offline evaluation methods, VSAS-Bench incorporates metrics for proactiveness and consistency, crucial for streaming VLMs. The benchmark includes over 18,000 temporally dense annotations across various domains and task types, along with standardized evaluation protocols and metrics to isolate specific streaming VLM capabilities. Their evaluations showed that adapted conventional VLMs can outperform specialized streaming models, with Qwen3-VL-4B achieving a 3% lead over the top-performing streaming VLM on their benchmark. AI