Research map / Vision-language and multimodal / Multimodal LLMs
Magic-MM-Embedding: Towards Visual-Token-Efficient Universal Multimodal Embedding with MLLMs
Qi Li, Yanzhe Zhao, Yongxin Zhou, Yameng Wang, Yandong Yang, Yuanjia Zhou, Jinxiang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add visual token compression (64 vs 256 tokens) (ablation) | InternVL3-2B embedder · MMEB+VisDoc average (warm-up) | average score: 66.1 -> 65.4 | No clear effect |
| add stage 1 capability restoration (ablation) | Magic-MM-Embedding 2B · MMEB (36 datasets) (1.9B) | Precision@1: 64.7 -> 65.4 | Improved |
| add stage 2 contrastive pretraining (ablation) | Magic-MM-Embedding 2B · MMEB (36 datasets) (1.9B) | Precision@1: 67.1 -> 68.0 | Improved |
| add stage 3 task-aware finetuning (ablation) | Magic-MM-Embedding 2B · MMEB (36 datasets) (1.9B) | Precision@1: 65.4 -> 68.0 | Improved |
| add reranker at inference (ablation) | Magic-MM-Embedding 2B · MMEB | Precision@1: 68.0 -> 70.2 | Improved |
| add reranker at inference (ablation) | Magic-MM-Embedding 2B · VisDoc | NDCG@5: 72.1 -> 73.3 | Improved |
| use MLLM-judged hard negatives instead of rule-based (ablation) | Magic-MM-Embedding 2B stage 3 · MMEB (36 datasets) (1.9B) | Precision@1: 67.0 -> 68.0 | Improved |
| swap judge Qwen3-VL-8B for InternVL3-8B (ablation) | Magic-MM-Embedding 2B stage 3 · MMEB (36 datasets) (1.9B) | Precision@1: 68.0 -> 67.8 | No clear effect |
| use bilinear interpolation instead of adaptive average pooling (ablation) | InternVL3-VTC 2B · VisDoc (short warm-up) | NDCG@5: 63.8 -> 64.6 | Improved |
| use bilinear interpolation instead of learnable Conv2d (ablation) | InternVL3-VTC 2B · MMEB (short warm-up) | Precision@1: 35.1 -> 57.4 | Improved |
| add visual token compression to cut training time (ablation) | InternVL3-2B · MMEB | Precision@1: 52h 43m 35s -> 24h 38m 28s | Improved |
| use LoRA rank 16 instead of 32 (ablation) | Magic-MM-Embedding 2B warm-up · MMEB+VisDoc average | average score: 65.1 -> 65.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- UniME-V2 — UniME-V2: MLLM-as-a-Judge for Universal Multimodal Embedding Learning (AAAI 2026) · better on 3 of 3 shared results
- ColPali — ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS (ICLR 2025 · similarity 0.93)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.91)
- Scaling Language-centric Omnimodal Representation Learning (NeurIPS 2025 · similarity 0.91)
- Bridging Modalities: Improving Universal Multimodal Retrieval by Multimodal Large Language Models (CVPR 2025 · similarity 0.91)
- Accelerating Multimodal Large Language Models by Searching Optimal Vision Token Reduction (CVPR 2025 · similarity 0.90)
- What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
