{"paper":{"title":"DocAtlas: Multilingual Document Understanding Across 80+ Languages","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Direct preference optimization with rendering-derived ground truth improves multilingual document understanding across 82 languages without base-language degradation.","cross_cats":["cs.CV","cs.LG"],"primary_cat":"cs.CL","authors_text":"Abdullah Sohail, Ahmed Heakl, Ahmed Nassar, Fahad Shahbaz Khan, Imran Razzak, Peter W. J. Staar, Rania Elbadry, Salman Khan, Youssef Mohamed","submitted_at":"2026-05-12T18:09:38Z","abstract_excerpt":"Multilingual document understanding remains limited for low-resource languages due to scarce training data and model-based annotation pipelines that perpetuate existing biases. We introduce DocAtlas, a framework that constructs high-fidelity OCR datasets and benchmarks covering 82 languages and 9 evaluation tasks. Our dual pipelines, differential rendering of native DOCX documents and synthetic LaTeX-based generation for right-to-left scripts produce precise structural annotations in a unified DocTag format encoding layout, text, and component types, without learned models for core annotation."},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Direct Preference Optimization (DPO) using rendering-derived ground truth as positive signal achieves stable multilingual adaptation, improving both in-domain (+1.9%) and out-of-domain (+1.8%) accuracy without measurable base-language degradation, where supervised fine-tuning degrades out-of-domain performance by up to 21%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The differential rendering of native DOCX documents and synthetic LaTeX-based generation produce precise structural annotations that accurately represent real-world document distributions across all 82 languages without introducing new biases or distribution shifts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DocAtlas creates multilingual document datasets across 82 languages and shows DPO with rendered ground truth improves model accuracy by 1.7-1.9% without degrading base-language performance, unlike supervised fine-tuning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Direct preference optimization with rendering-derived ground truth improves multilingual document understanding across 82 languages without base-language degradation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"38e7ff549c4cd97f2b79d1de71759e2cb5273a209f9b54a81f96523094dc1f0f"},"source":{"id":"2605.12623","kind":"arxiv","version":1},"verdict":{"id":"08b03706-e999-40f7-97fc-16d1abfacb79","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:58:47.305197Z","strongest_claim":"Direct Preference Optimization (DPO) using rendering-derived ground truth as positive signal achieves stable multilingual adaptation, improving both in-domain (+1.9%) and out-of-domain (+1.8%) accuracy without measurable base-language degradation, where supervised fine-tuning degrades out-of-domain performance by up to 21%.","one_line_summary":"DocAtlas creates multilingual document datasets across 82 languages and shows DPO with rendered ground truth improves model accuracy by 1.7-1.9% without degrading base-language performance, unlike supervised fine-tuning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The differential rendering of native DOCX documents and synthetic LaTeX-based generation produce precise structural annotations that accurately represent real-world document distributions across all 82 languages without introducing new biases or distribution shifts.","pith_extraction_headline":"Direct preference optimization with rendering-derived ground truth improves multilingual document understanding across 82 languages without base-language degradation."},"references":{"count":85,"sample":[{"doi":"","year":null,"title":"FirstName LastName , title =","work_id":"d9cab501-317f-4237-9e32-b5ead5964402","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"FirstName Alpher , title =","work_id":"42297990-8783-41a1-b0fa-8ccdbf630852","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Journal of Foo , volume = 13, number = 1, pages =","work_id":"65a8b3d0-af84-4f68-87eb-101c85ab18b2","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Journal of Foo , volume = 14, number = 1, pages =","work_id":"b3089947-bd36-4a24-9199-cc535e299537","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"FirstName Alpher and FirstName Gamow , title =","work_id":"caed320b-7cdc-41ca-bb08-00fb14feec62","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":85,"snapshot_sha256":"6c0e904e1734960766610409aa89dcf852a4115c9170d7aea2f25766027f1c43","internal_anchors":14},"formal_canon":{"evidence_count":2,"snapshot_sha256":"de605d39aaae8914b8445ffd563e90faab484cae48c46f137c0e77f8d585e400"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}