Research map / Vision-language and multimodal / CLIP and zero-shot
When Is Rank-1 Enough? Geometry-Guided Initialization for Parameter-Efficient Fine-Tuning
Haoran Zhao, Caren Han, Eduard Hovy
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 |
|---|---|---|---|
| initialize rank-1 LoRA direction with modality-gap vector | rank-1 LoRA, BLIP-2 · COCO Captioning | CIDEr: 98.08 -> 140.59 | Improved |
| initialize rank-1 LoRA with Gap-Init | rank-1 LoRA, BLIP-2 · VQAv2 | accuracy: 15.58 -> 57.23 | Improved |
| use Gap-Init instead of PiSSA at rank 1 | rank-1 LoRA · VQAv2 | accuracy: 63.35 -> 57.23 | Hurt |
| use Gap-Init instead of PiSSA at rank 1 | rank-1 LoRA · POPE | F1: 86.9 -> 92.4 | Improved |
| use Gap-Init instead of PiSSA at rank 1 | rank-1 LoRA · Flickr30k zero-shot | CIDEr: 76.00 -> 79.60 | Improved |
| add gap-guided GG-Safe layer selection with random init (ablation) | rank-1 LoRA · COCO Captioning | CIDEr: 98.08 -> 102.15 | Improved |
| use GG-Safe instead of naive layer selection with Gap-Init (ablation) | Gap-Init · COCO Captioning | CIDEr: 140.59 -> 140.06 | No clear effect |
| increase DoRA rank from 1 to 8 (ablation) | DoRA · COCO Captioning | CIDEr: 140.51 -> 139.55 | No clear effect |
| use Gap-Init instead of random init across 5 seeds (ablation) | rank-1 LoRA · COCO Captioning | CIDEr std: 7.10 -> 1.44 | Improved |
| use Gap-Init on Qwen2-VL-7B | rank-1 LoRA · COCO Captioning | CIDEr: 143.67 -> 144.11 | Improved |
| use Gap-Init on Gemma3-4B | rank-1 LoRA · COCO Captioning | CIDEr: 90.35 -> 95.24 | Improved |
| calibrate gap with out-of-domain data (ablation) | Gap-Init · calibration ablation | performance | Hurt |
Closest papers by meaning
- MARS: Harmonizing Multimodal Convergence via Adaptive Rank Search (NeurIPS 2026 · similarity 0.91)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.90)
- AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning (ICLR 2025 · similarity 0.88)
- GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (NeurIPS 2025 · similarity 0.88)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.88)
- HyperET: Efficient Training in Hyperbolic Space for Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
