{"paper":{"arxiv_id":"2006.03654","title":"DeBERTa: Decoding-enhanced BERT with Disentangled Attention","abstract":"Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. In this paper we propose a new model architecture DeBERTa (Decoding-enhanced BERT with disentangled attention) that improves the BERT and RoBERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions, respectively. Second, an enhanced mask decoder is used to incorporate absolute positions in the decoding layer to predict the masked tokens in model pre-training. In addition, a new virtual adversarial training method is used for fine-tuning to improve models' generalization. We show that these techniques significantly improve the efficiency of model pre-training and the performance of both natural language understanding (NLU) and natural language generation (NLG) downstream tasks.","primary_category":"cs.CL","venue":"ICLR 2021","published_at":null,"latest_version":1,"withdrawn":false},"latest_version":{"id":"69c7b1c4-1cdf-46c9-95d7-f4f48cc44a18","version":1,"source_url":"https://arxiv.org/abs/2006.03654","rendered_html_url":null,"rendering_engine":null},"verdict":{"id":"267f07ee-8cbf-4bf0-9cf3-29c1447feb76","kind":"POST","status":"reproduced","score":0.9169667867146859,"confidence":0.85,"agent_version":"v0.1.0-deberta-v2-mnli-microslice","computed_at":"2026-05-14T23:21:19.951Z","is_current":true,"claim_citation":{"paper_arxiv_id":"2006.03654","section":"Table 2","row":"DeBERTalarge","column":"MNLI-m/mm","reported_value":91.1,"reported_metric":"accuracy","quoted_text":"DeBERTalarge 91.1/91.1 95.5/90.1 90.7/88.0 86.8 91.4/91.0 90.8 93.8","pdf_page":7,"notes":"Table 2 row 'DeBERTa large' column 'MNLI-m/mm (Acc)'. The same model appears in Table 1 row 'DeBERTalarge' column 'MNLI-m/mm' with the identical value 91.1/91.1; both rows describe the 24-layer 1024-hidden DeBERTa-large model. The microsoft/deberta-large-mnli checkpoint is the public MNLI fine-tune of this exact model class. The 91.1 value also appears in the paper's abstract: 'MNLI by +0.9% (90.2% vs. 91.1%)'. The 2026-05-13 Phase B PDF-extractor upgrade (pdfjs-dist) now verifies this quote against the real arXiv PDF on the cited page (Table 2, page 7); see tests/unit/pdf-extractor-real-papers.test.ts."},"protocol_match":"exact"},"verdicts":{"post":{"id":"267f07ee-8cbf-4bf0-9cf3-29c1447feb76","kind":"POST","status":"reproduced","score":0.9169667867146859,"confidence":0.85,"agent_version":"v0.1.0-deberta-v2-mnli-microslice","computed_at":"2026-05-14T23:21:19.951Z","is_current":true,"claim_citation":{"paper_arxiv_id":"2006.03654","section":"Table 2","row":"DeBERTalarge","column":"MNLI-m/mm","reported_value":91.1,"reported_metric":"accuracy","quoted_text":"DeBERTalarge 91.1/91.1 95.5/90.1 90.7/88.0 86.8 91.4/91.0 90.8 93.8","pdf_page":7,"notes":"Table 2 row 'DeBERTa large' column 'MNLI-m/mm (Acc)'. The same model appears in Table 1 row 'DeBERTalarge' column 'MNLI-m/mm' with the identical value 91.1/91.1; both rows describe the 24-layer 1024-hidden DeBERTa-large model. The microsoft/deberta-large-mnli checkpoint is the public MNLI fine-tune of this exact model class. The 91.1 value also appears in the paper's abstract: 'MNLI by +0.9% (90.2% vs. 91.1%)'. The 2026-05-13 Phase B PDF-extractor upgrade (pdfjs-dist) now verifies this quote against the real arXiv PDF on the cited page (Table 2, page 7); see tests/unit/pdf-extractor-real-papers.test.ts."},"protocol_match":"exact"},"pre":null}}