Research map / Continual learning / Continual fine-tuning of LLMs
CoPE: Continual Probe-guided Expansion for Large Vision-Language Models
Ziqin Wang, Hengyuan Zhao, Qixin Sun, Kaiyou Song, Yilin Li, Xiaolin Hu, Qingpei Guo, Linjiang Huang, Si Liu
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 expert expansion with frozen old experts (PGKE) (ablation) | CoPE · CoIN Last | BWT: -17.68 -> -3.58 | Improved |
| probe-selected layers instead of random layers for expansion (ablation) | CoPE · CoIN Immediate | accuracy: 76.73 -> 79.01 (SQA) | Improved |
| probe-selected layers instead of expanding every layer (ablation) | CoPE · CoIN Immediate | accuracy / parameters: 0.84 param-ratio on OCR-VQA, 62.01 vs 59.78 | No clear effect |
| add Probabilistic Task Locator instead of last-task router (ablation) | CoPE · CoIN Last | accuracy: 29.68 -> 94.50 (ImageNet) | Improved |
| initialize new experts from the nearest expert instead of zero init (ablation) | CoPE · CoIN Immediate | mean accuracy: 60.77 -> 62.81 | Improved |
| add load-balancing loss in probe stage (ablation) | CoPE · CoIN Immediate | experts added: 24.25 -> 22.75 | Improved |
| increase added experts per layer from 1 to 8 (ablation) | CoPE · CoIN Ref | accuracy: 25.63 -> 41.51 | Improved |
| increase added experts per layer from 1 to 8 (ablation) | CoPE · ScienceQA | accuracy: 79.01 -> 82.01 | Improved |
| change task order (reverse or random) (ablation) | CoPE · CoIN Last | BWT: -3.58 vs -3.66 / -3.51 | No clear effect |
| average-pooled all-layer features for PTL instead of last-layer feature (ablation) | CoPE · CoIN Last | mean accuracy: 59.23 -> 55.34 | Hurt |
| use stricter selection threshold alpha 0.4 instead of 0.8 (ablation) | CoPE · CoIN Immediate | avg accuracy: 62.81 -> 61.72 | Hurt |
| use looser threshold alpha 1.2 instead of 0.8 (ablation) | CoPE · CoIN Immediate | avg params: 0.71 -> 0.84 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SEFE — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scalable Language Model with Generalized Continual Learning (ICLR 2024 · similarity 0.89)
- Overcoming Generic Knowledge Loss with Selective Parameter Update (CVPR 2024 · similarity 0.89)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.89)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.89)
- Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters (CVPR 2024 · similarity 0.89)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 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.
