Research map / Vision-language and multimodal / Multimodal LLMs
Are They the Same? Exploring Visual Correspondence Shortcomings of Multimodal LLMs
Yikang Zhou, Tao Zhang, Shilin Xu, Shihao Chen, Qianyu Zhou, Yunhai Tong, Shunping Ji, Jiangning Zhang, Lu Qi, Xiangtai Li
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 MMVM SFT data with LLaVA SFT data (ablation) | InternVL2-4B · MMVM benchmark (ablation, 30% data) | Overall accuracy: 17.62 -> 32.38 | Improved |
| add object-level contrastive learning to the MLLM visual encoder only (ablation) | InternVL2-4B + Data · MMVM benchmark (ablation, 30% data) | Overall accuracy: 32.38 -> 34.05 | Improved |
| add RADIO vision expert without contrastive alignment (ablation) | InternVL2-4B + Data · MMVM benchmark (ablation, 30% data) | Overall accuracy: 32.38 -> 32.25 | No clear effect |
| add vision expert with OCL pre-training of its adapter (ablation) | InternVL2-4B + Data + OCL · MMVM benchmark (ablation, 30% data) | Overall accuracy: 34.05 -> 40.45 | Improved |
| add instruction augmentation with object-level representations (ablation) | InternVL2-4B + Data + OCL + VE · MMVM benchmark (ablation, 30% data) | Overall accuracy: 40.45 -> 45.83 | Improved |
| pre-train adapter with region-region contrastive loss instead of image-text (ablation) | CoLVA · MMVM benchmark (adapter pre-training ablation) | Overall accuracy: 33.64 -> 40.45 | Improved |
| pre-train adapter with region-region contrastive loss instead of region-text (ablation) | CoLVA · MMVM benchmark (adapter pre-training ablation) | Overall accuracy: 30.93 -> 40.45 | Improved |
| use RADIO instead of DINOv2 as vision expert (ablation) | CoLVA · MMVM benchmark (vision expert alternatives) | Overall accuracy: 40.34 -> 45.83 | Improved |
| use RADIO instead of ConvNext-L as vision expert (ablation) | CoLVA · MMVM benchmark (vision expert alternatives) | Overall accuracy: 39.80 -> 45.83 | Improved |
| integrate CoLVA into LLaVA1.5-7B without multi-image training | LLaVA1.5-7B · MMVM benchmark (generalization to other MLLMs) | Overall accuracy: 14.64 -> 36.56 | Improved |
| integrate CoLVA into Qwen2VL-2B | Qwen2VL-2B · MMVM benchmark (generalization to other MLLMs) | Overall accuracy: 15.69 -> 47.48 | Improved |
| integrate CoLVA into InternVL2-4B on MMMU Val | InternVL2-4B · MMMU Val | Overall accuracy: 47.56 -> 44.11 | Hurt |
Closest papers by meaning
- Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs (CVPR 2024 · similarity 0.92)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.90)
- Visual Perception by Large Language Model’s Weights (NeurIPS 2024 · similarity 0.90)
- MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models (ICLR 2025 · similarity 0.90)
- Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Models (CVPR 2025 · similarity 0.90)
- MLLM-CompBench: A Comparative Reasoning Benchmark for Multimodal LLMs (NeurIPS 2024 · similarity 0.89)
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.
