Research map / Vision-language and multimodal / CLIP and zero-shot
Enhanced Continual Learning of Vision-Language Models with Model Fusion
Haoyuan Gao, Zicong Zhang, Yuqi Wei, Linglan Zhao, Guilin Li, Yexin Li, Bo Wang, Linghe Kong, Weiran Huang
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 rescalers to the task triggers (ablation) | ConDU (FT) · MTIL (order I) | Average accuracy (%): 74.2 -> 78.8 | Improved |
| add rescalers to the task triggers (ablation) | ConDU (FT) · MTIL (order I) | Last accuracy (%): 77.1 -> 87.1 | Improved |
| add rescalers to the task triggers (ablation) | ConDU (FT) · MTIL (order I) | Transfer accuracy (%): 70.4 -> 70.8 | Improved |
| compute prototype similarity with pre-trained model features instead of task-expert features (ablation) | ConDU (FT) · MTIL (order I) | Transfer accuracy (%): 67.7 -> 70.8 | Improved |
| compute prototype similarity with pre-trained model features instead of task-expert features (ablation) | ConDU (FT) · MTIL (order I) | Average accuracy (%): 77.8 -> 78.8 | Improved |
| aggregate K=5 task experts instead of K=1 (ablation) | ConDU (FT) · MTIL (order I) | Transfer accuracy (%): 69.5 -> 70.8 | Improved |
| increase K from 5 to 11 (ablation) | ConDU (FT) · MTIL (order I) | Transfer accuracy (%): 70.8 -> 70.2 | Hurt |
| aggregate decoupled task experts instead of using the unified model directly (ablation) | ConDU (FT) · MTIL (order I) | Transfer accuracy (%): 47.1 -> 70.8 | Improved |
| run on RTX 4090 instead of Ascend 910B | ConDU (FT) · MTIL (order I) | Transfer accuracy (%): 70.2 -> 70.8 | No clear effect |
| use LoRA instead of full fine-tuning | ConDU · Few-shot MTIL | Average accuracy (%): 72.3 -> 72.7 | Improved |
| use LoRA instead of full fine-tuning | ConDU · TinyImageNet class-incremental (10 steps) | accuracy (%): 71.74 -> 68.80 | Hurt |
| use ConDU instead of the best baseline (DPeCLIP) | ConDU (FT) · Task-agnostic MTIL | Average accuracy (%): 76.1 -> 78.1 | Improved |
Closest papers by meaning
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (CVPR 2024 · similarity 0.92)
- 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)
- Decomposing and Composing: Towards Efficient Vision-Language Continual Learning via Rank-1 Expert Pool in a Single LoRA (AAAI 2026 · similarity 0.91)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.90)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · 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.
