Different Strokes for Different Folks: Writer Identification for Historical Arabic Manuscripts

The more significant contribution here may be the labeling work itself: manually verifying and expanding writer attribution from 28% to 87% coverage across nearly 19,000 line images is a data curation effort that makes future benchmarking possible at all, not just a preprocessing step. Without that groundwork, any model results would be nearly uninterpretable.

This is largely disconnected from recent activity covered on Modelwire, which has focused on LLM coding tools, agentic platforms, and commercial AI deployment. The closest thematic thread is the reliability-of-evaluation problem: the April 16 arXiv paper on LLM judge reliability ('Diagnosing LLM Judge Reliability') raised concerns about how confidently we can trust automated assessments, and the manuscript work surfaces a parallel issue in a very different domain, specifically that you cannot evaluate writer identification models responsibly without first trusting your ground-truth labels. Both papers are, at root, about the conditions required for valid measurement.

Watch whether the Muharaf dataset with its expanded labels gets adopted by other research groups within the next 12 months. If it does, the baseline numbers published here will face direct challenge and either hold or reveal the limits of the initial methodology.