Reproduce “MIBench: Evaluating Multimodal Large Language Models over Multiple Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang and 3 more
MIBench evaluates whether multimodal models can handle multiple images alongside text, rather than relying on single-image performance. It covers 13 tasks across instruction, knowledge-seeking, and in-context-learning scenarios, using multiple-choice accuracy and short-answer exact matching. On Temporal Reasoning, GPT-4o reaches 68.0% accuracy and MMICL 47.0%; these results cover tested models and inputs of 2–8 images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.