Research map / Vision-language and multimodal / Multimodal LLMs
Deciphering Cross-Modal Alignment in Large Vision-Language Models via Modality Integration Rate
Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jiaqi Wang, Weiming Zhang, Nenghai Yu
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 MoCa learnable scale-and-shift on vision tokens | LLaVA-v1.5 7B · 7-9 benchmark average | Average score: 59.1 -> 60.6 | Improved |
| add MoCa | LLaVA-v1.5 7B · MIR | MIR: 3.374 -> 3.162 | Improved |
| add MoCa | Mini-Gemini 7B · benchmark average | Average score: 62.1 -> 63.0 | Improved |
| add MoCa | Mini-Gemini 7B · MIR | MIR: 2.667 -> 2.514 | Improved |
| lengthen pre-training captions from 15.2 to 181.2 tokens (ablation) | LLaVA-v1.5 7B · post-SFT average | Average score: 63.8 -> 64.4 | Improved |
| lengthen captions from 127.1 to 181.2 (ablation) | LLaVA-v1.5 7B · SEED-Img | accuracy: 67.6 -> 67.6 | No clear effect |
| lower learning rate from 3e-3 to 3e-4 (ablation) | LLaVA-v1.5 7B pre-training · post-SFT average | Average score: 58.7 -> 60.0 | Improved |
| increase warmup ratio from 3e-2 to 5e-2 (ablation) | LLaVA-v1.5 7B pre-training · post-SFT average | Average score: 59.5 -> 59.6 | No clear effect |
| use linear instead of cosine LR scheduler (ablation) | LLaVA-v1.5 7B pre-training · post-SFT average | Average score: 59.5 -> 59.5 | No clear effect |
| unlock entire LLM during pre-training instead of projector only (ablation) | LLaVA-v1.5 7B, ~1.9M data · post-SFT average | Average score: 64.0 -> 65.9 | Improved |
| unlock former half of LLM instead of projector only (ablation) | LLaVA-v1.5 7B, ~1.9M data · post-SFT average | Average score: 64.0 -> 65.9 | Improved |
| unlock LoRA on LLM with merge instead of projector only (ablation) | LLaVA-v1.5 7B, ~1.9M data · post-SFT average | Average score: 64.0 -> 64.5 | Improved |
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.
- Mini-Gemini — Mini-Gemini: Mining the Potential of Multi-Modality Vision Language Models (TPAMI 2025) · better on 2 of 2 shared results
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
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 · similarity 0.94)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.93)
- DeepAlign: Mitigating Modality Conflict through Modality-Specific Alignment (CVPR 2026 · similarity 0.92)
- Data Selection for Fine-tuning Vision Language Models via Cross Modal Alignment Trajectories (ICML 2026 · similarity 0.91)
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.91)
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.91)
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.
