Research map / Vision-language and multimodal / Multimodal LLMs
Dive into the implicit biases of low-rank vision-language alignment
Mingjia Shi, Shuo Wang, Xiaobo Wang, Sifan Zhou, Kai Wang, Tianyu Fu, Chenxu Zhao, Anyang Su, Ping Jiang, Minghui Wu
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 low-rank instead of full-parameter LLM updates in alignment (ablation) | VLM alignment (1.4B-13B) · MME-perception | MME-P: 236.6 -> 1079.7 (13B) | Improved |
| use low-rank instead of full-parameter alignment (ablation) | VLM alignment 13B · Training time | hours: 35.32 -> 13.23 | Improved |
| use low-rank alignment on Qwen3-8B (ablation) | Qwen3-8B · MMBench | accuracy: 18.1 -> 48.7 | Improved |
| use low-rank alignment on Qwen3-8B (ablation) | Qwen3-8B · POPE | POPE: 92.7 -> 25.1 | Hurt |
| use low-rank alignment on Qwen3-8B (ablation) | Qwen3-8B · MMMU-Pro | accuracy: 12.1 -> 5.7 | Hurt |
| use LoHa instead of LoRA as low-rank operator (ablation) | low-rank alignment · MME-perception (Vicuna-13B) | MME-P: 375.9 -> 1079.7 | Improved |
| use LoKr at 13B instead of LoHa (ablation) | low-rank alignment · Table 5 13B POPE | POPE: 67.4 -> 37.6 | Hurt |
| freeze vision encoder during alignment (ablation) | LoRA alignment (CLIP) · GQA | accuracy: 60.18 -> 62.47 | Improved |
| apply LoRA to both attention and MLP instead of one (ablation) | LoRA alignment · MME-P | MME-P: 1384.71 -> 1511.61 (vs MLP-only) | Improved |
| use low-rank instead of full with tuned weight decay (ablation) | alignment · six benchmarks | accuracy: 1432.68 -> 1511.61 (MME-P) | Improved |
| use low-rank alignment with DINOv2 or SigLIPv2 encoder (ablation) | alignment · MME-P / GQA | MME-P: 1503.69 -> 1481.45 (DINOv2 MME-P); 1406.61 -> 1474.70 (SigLIPv2) | Mixed |
| increase LoRA rank (ablation) | low-rank alignment · knowledge benchmarks (GQA, SQA, TextVQA) | accuracy | No clear effect |
Closest papers by meaning
- Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment (ICLR 2025 · similarity 0.91)
- Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models (NeurIPS 2025 · similarity 0.90)
- Assessing and Learning Alignment of Unimodal Vision and Language Models (CVPR 2025 · similarity 0.90)
- LVLM-Aided Alignment of Task-Specific Vision Models (CVPR 2026 · similarity 0.90)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 2025 · similarity 0.90)
- Locality Alignment Improves Vision-Language Models (ICLR 2025 · 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.
