MMLongBench

eval
University of Edinburgh

Long-context vision-language benchmark (NeurIPS 2025 spotlight; repo EdinburghNLP/MMLongBench): 13,331 examples across 5 task categories at 8K–128K context lengths, with 46 models evaluated at release. An adopted benchmark for long-context VLM evaluation, from Pasquale Minervini's group with external coauthors; Edinburgh-hosted.

Paper (arXiv)GitHub

Evaluation Details

Questions 13,331
Domains 2
Scoring task accuracy
Domains: long-context, vision-language
evalmultimodalevaluation