Research map / Architectures and efficiency / Pruning and efficient tuning
Curvature-Guided Mixing for MLLM Adaptation
Jinglong Yang, Jiaxuan He, Wenjian Huang, Zhan Zhuang, Jianguo Zhang
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 pre-train curvature term and fine-tune curvature to magnitude score (full CGM† score) (ablation) | CGM† · LLaVA-1.5-7B OKVQA | Hscore: 54.0 -> 62.8 | Improved |
| add pre-train curvature term to magnitude score (ablation) | CGM† score · LLaVA-1.5-7B OKVQA | Hscore: 54.0 -> 61.7 | Improved |
| add fine-tune curvature to magnitude score (ablation) | CGM† score · LLaVA-1.5-7B OKVQA | Hscore: 54.0 -> 53.1 | No clear effect |
| add fine-tune curvature to pre-train+magnitude score (full) (ablation) | CGM† score · LLaVA-1.5-7B OKVQA | Hscore: 61.7 -> 62.8 | Improved |
| use full curvature score instead of magnitude-only (ablation) | CGM† · LLaVA-1.5-7B Flickr30k | Hscore: 53.59 -> 56.22 | Improved |
| use full curvature score instead of magnitude-only (ablation) | CGM† · Qwen-2.5VL-3B LaTeX-OCR | Hscore: 67.6 -> 72.7 | Improved |
| add fine-tune curvature to pre-train+magnitude score (full) (ablation) | CGM† · Qwen-2.5VL-3B LaTeX-OCR | Hscore: 71.7 -> 72.7 | Improved |
| vary sparsity ratio K from 10% to 90% (ablation) | CGM† · Qwen-2.5VL-3B (LaTeX-OCR, Flickr30k) | Pre-Avg | No clear effect |
| use small K (10%) (ablation) | CGM† · Qwen-2.5VL-3B | Hscore | Improved |
| use small alpha (0.1 to 0.15) (ablation) | CGM† · Qwen-2.5VL-3B | Hscore | Improved |
| CGM† vs fine-tuned model | CGM† · LLaVA-1.5-7B OKVQA | Pre-Avg: 50.4 -> 65.7 | Improved |
| CGM† vs Tailor | CGM† · LLaVA-1.5-7B OKVQA | Hscore: 59.6 -> 62.8 | Improved |
Closest papers by meaning
- OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging (ICLR 2026 · similarity 0.87)
- FourierMoE: Fourier Mixture-of-Experts Adaptation of Large Language Models (NeurIPS 2026 · similarity 0.86)
- Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer (ICLR 2025 · similarity 0.86)
- LoRASculpt: Sculpting LoRA for Harmonizing General and Specialized Knowledge in Multimodal Large Language Models (CVPR 2025 · similarity 0.86)
- On Path to Multimodal Generalist: General-Level and General-Bench (ICML 2025 · similarity 0.86)
- $\gamma-$MoD: Exploring Mixture-of-Depth Adaptation for Multimodal Large Language Models (ICLR 2025 · similarity 0.86)
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.
