Research map / Vision-language and multimodal / Multimodal LLMs
VersaViT: Enhancing MLLM Vision Backbones via Task-Guided Optimization
Yikun Liu, Yuan Liu, Shangzhe Di, Haicheng Wang, Zhongyin Zhao, Le Tian, Zhou Xiao, Jie Zhou, Jiangchao Yao, Yanfeng Wang, Weidi Xie
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 |
|---|---|---|---|
| apply VersaViT multi-task post-training | Qwen2-VL-ViT · OpenCompass average | average score: 64.8 -> 66.4 | Improved |
| apply VersaViT multi-task post-training | Qwen2-VL-ViT · ADE20k linear probe | mIoU: 33.6 -> 49.6 | Improved |
| apply VersaViT multi-task post-training | Qwen2-VL-ViT · KITTI linear probe | RMSE: 3.735 -> 3.136 | Improved |
| add VQA and captioning training only (ablation) | Qwen2-VL-ViT · ADE20k linear probe | mIoU: 33.6 -> 32.0 | Hurt |
| add depth task to VQA and captioning (ablation) | VersaViT · NYUv2 linear probe | RMSE: 0.557 -> 0.495 | Improved |
| add segmentation task to VQA, captioning and depth (ablation) | VersaViT · ADE20k linear probe | mIoU: 35.1 -> 49.6 | Improved |
| add segmentation task to VQA, captioning and depth (ablation) | VersaViT · OpenCompass average | average score: 66.3 -> 66.4 | Improved |
| use balanced loss weights 0.33 each instead of 0.25/0.50/0.25 (ablation) | VersaViT · OpenCompass average | average score: 65.7 -> 66.4 | Improved |
| apply multi-task post-training to Qwen2.5-VL-ViT (ablation) | Qwen2.5-VL-ViT · ADE20k linear probe | mIoU: 37.2 -> 47.2 | Improved |
| apply multi-task post-training to Siglip2 (ablation) | Siglip2-so400M-512 · OpenCompass average | average score: 59.3 -> 60.4 | Improved |
| use multi-task framework instead of more captioning/VQA data (35M) (ablation) | Qwen2-VL-ViT · ADE20k linear probe | mIoU: 34.7 -> 49.6 | Improved |
| apply multi-task post-training | Qwen2-VL-ViT · SPair semantic correspondence | recall: 17.05 -> 26.99 | Improved |
Closest papers by meaning
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.90)
- VisionLLM v2: An End-to-End Generalist Multimodal Large Language Model for Hundreds of Vision-Language Tasks (NeurIPS 2024 · similarity 0.90)
- Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment (CVPR 2025 · similarity 0.90)
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · similarity 0.90)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.90)
- Locality Alignment Improves Vision-Language Models (ICLR 2025 · 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.
