Research map / Vision-language and multimodal / Cross-modal learning
Multi-Modal Multi-Task Unified Embedding Model (M3T-UEM): A Task-Adaptive Representation Learning Framework
Rohan Sharma, Changyou Chen, Feng-Ju Chang, Seongjun Yun, Xiaohu Xie, Rui Meng, Dehong Xu, Alejandro Mottini, Qingjun Cui
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 |
|---|---|---|---|
| use task-aware contrastive loss instead of standard InfoNCE (ablation) | M3T-UEM · M-BEIR average | Recall: 52.7 -> 53.9 | Improved |
| add task-aware loss (5k-step ablation) (ablation) | M3T-UEM · M-BEIR average (5k-step) | retrieval average: 37.4 -> 38.0 | Improved |
| add two-stage training (alignment then refinement) at equal FLOPs (ablation) | M3T-UEM · M-BEIR average (5k-step) | retrieval average: 35.7 -> 38.0 | Improved |
| use 16 EOS tokens instead of one (ablation) | M3T-UEM · M-BEIR average (5k-step) | retrieval average: 37.6 -> 38.0 | Improved |
| add auxiliary language-modeling loss (ablation) | M3T-UEM · M-BEIR average (5k-step) | retrieval average: 37.9 -> 38.0 | No clear effect |
| use M3T-UEM LLM backbone versus OpenCLIP ViT-g-14 on multilingual retrieval | M3T-UEM · Multilingual zero-shot retrieval | Recall@5 text-to-image: 41.4 -> 72.9 | Improved |
| use M3T-UEM versus OpenCLIP ViT-g-14 on SugarCrepe Replace | M3T-UEM · SugarCrepe Replace | image-to-text accuracy: 81.7 -> 88.9 | Improved |
| use M3T-UEM versus MM-Embed on M-BEIR | M3T-UEM · M-BEIR average | Recall: 52.7 -> 53.9 | Improved |
| use M3T-UEM versus ViT-g-14 on throughput | M3T-UEM · CIFAR-100 throughput | samples/sec: 20.1 -> 18.9 | Hurt |
| use M3T-UEM versus ViT-g-14 on memory | M3T-UEM · inference memory | MB: 8,573 -> 21,778 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MM-Embed — MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS (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
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.90)
- Understanding Retrieval-Augmented Task Adaptation for Vision-Language Models (ICML 2024 · similarity 0.89)
- MmAP: Multi-Modal Alignment Prompt for Cross-Domain Multi-Task Learning (AAAI 2024 · similarity 0.88)
- Bridging Vision and Language Spaces with Assignment Prediction (ICLR 2024 · similarity 0.88)
- Beyond Global Similarity: Multi-Conditional Retrieval for Fine-Grained Cross-Modal Understanding (CVPR 2026 · similarity 0.87)
- Towards General Continuous Memory for Vision-Language Models (NeurIPS 2025 · similarity 0.87)
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.
