Research map / Continual learning / Continual fine-tuning of LLMs
Dynamic Multi-Layer Null Space Projection for Vision-Language Continual Learning
Borui Kang, Lei Wang, Zhiping Wu, Tao Feng, Yawen Li, Yang Gao, Wenbin Li
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 |
|---|---|---|---|
| apply multi-layer null space projection to vision branch only with dynamic lambda (ablation) | Adapter-based CLIP · TinyImageNet CIL (5 subsets) | Last accuracy: 78.24 -> 79.14 | Improved |
| apply multi-layer projection to both modalities (ablation) | Adapter-based CLIP · TinyImageNet CIL (5 subsets) | Last accuracy: 78.24 -> 78.12 | No clear effect |
| apply DMNSP to language branch only (ablation) | Adapter-based CLIP · TinyImageNet CIL (5 subsets) | Last accuracy: 78.24 -> 77.22 | Hurt |
| apply single-layer projection to vision only plus lambda (ablation) | Adapter-based CLIP · TinyImageNet CIL (5 subsets) | Last accuracy: 78.24 -> 78.46 | No clear effect |
| add traditional single-layer null space projection (GPM) to adapter CLIP | Adapter-based CLIP · TinyImageNet CIL (10 subsets) | Last forgetting rate: 8.75 -> 8.69 | No clear effect |
| add DMNSP to adapter CLIP | Adapter-based CLIP · TinyImageNet CIL (10 subsets) | Last forgetting rate: 8.75 -> 4.84 | Improved |
| add DMNSP to MoE4Adapters | MoE4Adapters · TinyImageNet CIL (5 subsets) | Average accuracy: 81.12 -> 81.19 | Improved |
| use simpler adapter with DMNSP instead of MoE4Adapters | CLIP continual learning · CIFAR100 training cost | time per iteration: 1.58s/it -> 0.23s/it | Improved |
| DMNSP vs TRGP on CIFAR100 50 subsets (average) | Adapter-based CLIP · CIFAR100 CIL (50 subsets) | Average accuracy: 83.70 -> 83.66 | No clear effect |
Closest papers by meaning
- Learning Without Forgetting for Vision-Language Models (TPAMI 2025 · similarity 0.92)
- Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models (ICLR 2026 · similarity 0.91)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.89)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.88)
- Visual Prompt Tuning in Null Space for Continual Learning (NeurIPS 2024 · similarity 0.88)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.88)
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.
