MMEB

MMEB (Massive Multimodal Embedding Benchmark) is a comprehensive benchmark that includes 4 meta-tasks and 36 datasets designed to evaluate universal multimodal embedding models across various downstream tasks.

Papers using MMEB (12)

LLaVE: Large Language and Vision Embedding Models with Hardness-Weighted Contrastive Learning2025 · 52 citesCombating Visual Neglect and Semantic Drift in Large Multimodal Models for Enhanced Cross-Modal Retrieval2026Adaptive Global and Fine-Grained Perceptual Fusion for MLLM Embeddings Compatible with Hard Negative Amplification2026CREM: Compression-Driven Representation Enhancement for Multimodal Retrieval and Comprehension2026Compressing then Matching: An Efficient Pre-training Paradigm for Multimodal Embedding2025FreeRet: MLLMs as Training-Free Retrievers2025VLM2Vec-V2: Advancing Multimodal Embedding for Videos, Images, and Visual Documents2025Improve Multi-Modal Embedding Learning via Explicit Hard Negative Gradient Amplifying2025MoCa: Modality-aware Continual Pre-training Makes Better Bidirectional Multimodal Embeddings2025Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLMs2025mmE5: Improving Multimodal Multilingual Embeddings via High-quality Synthetic Data2025VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks2024