ViDoRe v3.1 Reranking

Cite this benchmark

@article{loison2026vidorev3comprehensiveevaluation,
  archiveprefix = {arXiv},
  author = {António Loison and Quentin Macé and Antoine Edy and Victor Xing and Tom Balough and Gabriel Moreira and Bo Liu and Manuel Faysse and Céline Hudelot and Gautier Viaud},
  eprint = {2601.08620},
  primaryclass = {cs.AI},
  title = {ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios},
  url = {https://arxiv.org/abs/2601.08620},
  year = {2026},
}

Reranking of visual documents on the ViDoRe v3.1 domains. For each domain, the candidates are the top-50 pages retrieved by one of four first-stage retrievers; BM25, bge-m3, Qwen3-VL-Embedding-2B and Nemotron-3.5-Embed-8B, so rerankers are compared on candidate lists from various model types. Includes both open and closed datasets; to submit results on private tasks, please open an issue.

Reference paper →

Languages 6

Tasks 40

Task Types 1

Models 0