{
  "id": "2509.25085",
  "version": "v2",
  "published": "2025-09-29T17:23:54.000Z",
  "updated": "2025-10-01T17:49:17.000Z",
  "title": "jina-reranker-v3: Last but Not Late Interaction for Document Reranking",
  "authors": [
    "Feng Wang",
    "Yuqing Li",
    "Han Xiao"
  ],
  "comment": "early draft, CodeIR table needs to be updated (qwen baselines are missing)",
  "categories": [
    "cs.CL",
    "cs.AI",
    "cs.IR"
  ],
  "abstract": "jina-reranker-v3 is a 0.6B parameter multilingual document reranker that introduces a novel last but not late interaction. Unlike late interaction models such as ColBERT that perform separate encoding followed by multi-vector matching, our approach conducts causal self-attention between query and documents within the same context window, enabling rich cross-document interactions before extracting contextual embeddings from the last token of each document. This compact architecture achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being significant smaller than generative listwise rerankers.",
  "revisions": [
    {
      "version": "v2",
      "updated": "2025-10-01T17:49:17.000Z"
    }
  ],
  "analyses": {
    "subjects": [
      "68T50",
      "I.2.7"
    ],
    "keywords": [
      "late interaction",
      "document reranking",
      "6b parameter multilingual document reranker",
      "compact architecture achieves state-of-the-art beir",
      "architecture achieves state-of-the-art beir performance"
    ],
    "note": {
      "typesetting": "TeX",
      "pages": 0,
      "language": "en",
      "license": "arXiv",
      "status": "editable"
    }
  }
}