arXiv cs.CL
·
18 hours ago
Researchers developed WikiSTAR, a system using language models to identify and categorize scientifically meaningful edits in Wikipedia's revision history. The system applies an LLM classifier with a multi-label taxonomy to tag edit types including technical terms, research findings, and narrative changes. Expert users found the system revealed new patterns in how scientific knowledge evolves on Wikipedia that were previously hidden by routine edits.
arXiv cs.AI
·
18 hours ago
Researchers developed a multi-model summarization framework that uses multiple fine-tuned transformer models to generate candidate summaries, then selects the best one based on automatic evaluation metrics. The framework achieved a BERTScore of 88.63% on the CNN/DailyMail dataset and outperformed LLMs including GPT3-D2, Falcon-7b, and Mpt-7b. This approach improves consistency by avoiding reliance on single models that perform unevenly across different article types.
arXiv cs.AI
·
18 hours ago
Researchers developed a multi-expert routing system for Manchu historical document OCR that uses a lightweight classifier to direct different document styles to specialized models trained through iterative fine-tuning. The routed system achieved character error rates of 0.30 percent on regular script, 1.57 percent on memorials, and 4.83 percent on running script, matching performance of domain-specific specialists selected for each style. This approach enables effective OCR for low-resource historical documents with multiple visual styles by reusing checkpoints as domain experts rather than training separate models from scratch.
arXiv cs.AI
·
18 hours ago
Researchers introduced OvisOCR2, a 0.8 billion parameter document parsing model that converts document images into Markdown format including text, formulas, tables, and visual regions. The model achieved a score of 96.58 on OmniDocBench v1.6 and 75.06 on PureDocBench, surpassing previous pipeline-based methods on these leaderboards. The end-to-end approach enables better generalization across diverse document types and challenging scenarios compared to existing methods.
