Research map / Vision-language and multimodal / Multimodal LLMs
From One-to-One to Many-to-Many: Dynamic Cross-Layer Injection for Deep Vision-Language Fusion
Cheng Chen, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, Peng Di, Hang Yu, Lianli Gao
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 AMP only (LoRA multi-projection) (ablation) | LLaVA-OV-0.5B · 9-benchmark ablation sum | sum score: 366.72 -> 367.89 | No clear effect |
| add AGF gating alone (ablation) | LLaVA-OV-0.5B · 9-benchmark ablation sum | sum score: 366.72 -> 370.60 | Improved |
| add AMP + AGF (CLI) (ablation) | LLaVA-OV-0.5B · 9-benchmark ablation sum | sum score: 366.72 -> 371.51 | Improved |
| fully fine-tune the projector in place of LoRA AMP, with AGF (ablation) | LLaVA-OV-0.5B · 9-benchmark ablation sum | sum score: 371.51 -> 407.01 | Improved |
| add CLI to the baseline projector | LLaVA-OV-7B · 9 doc/perception benchmarks | sum score: 650.9 -> 660.6 | Improved |
| add CLI to the baseline projector | LLaVA-OV-0.5B · 9 doc/perception benchmarks | sum score: 541.2 -> 542.9 | Improved |
| add CLI to the baseline projector | LLaVA-OV-7B · 9 reasoning benchmarks | sum score: 556.2 -> 559.4 | Improved |
| add CLI to LLaVA-1.5-7B | LLaVA-1.5-7B · 9 reasoning benchmarks | partial sum: 433.8 -> 442.4 | Improved |
| add DeepStack to the baseline | LLaVA-OV-7B · 9 doc/perception benchmarks | sum score: 650.9 -> 611.8 | Hurt |
| add Shallow-Layer Injection (SLI) | LLaVA-OV-7B · 9 reasoning benchmarks | sum score: 556.2 -> 551.1 | Hurt |
| add CLI | LLaVA-OV-7B · LLaVA-in-the-Wild | score: 68.0 -> 74.5 | Improved |
| replace CLI by equal-parameter static one-to-one Parallel injection (ablation) | LLaVA-OV-7B · LLaVA-in-the-Wild | score: dropping 6.5 points | Hurt |
Closest papers by meaning
- VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models (CVPR 2025 · similarity 0.90)
- From Pixels to Words -- Towards Native Vision-Language Primitives at Scale (ICLR 2026 · similarity 0.89)
- Florence-VL: Enhancing Vision-Language Models with Generative Vision Encoder and Depth-Breadth Fusion (CVPR 2025 · similarity 0.89)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.89)
- CoVLM: Composing Visual Entities and Relationships in Large Language Models Via Communicative Decoding (ICLR 2024 · similarity 0.89)
- Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference (CVPR 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.
