Research map / Continual learning / Online and replay-based
PromptFusion: Decoupling Stability and Plasticity for Continual Learning
Haoran Chen, Zuxuan Wu, Xintong Han, Menglin Jia, Yu-Gang Jiang
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 the weight mask W balancing old and new classes (ablation) | PromptFusion · Split-CIFAR100 | average accuracy: 84.2 -> 88.5 | Improved |
| add learnable fusion weight lambda instead of simple summation (ablation) | PromptFusion · Split-CIFAR100 | average accuracy: 87.4 -> 88.5 | Improved |
| add image prompts to augment CoOp (ablation) | PromptFusion · Split-CIFAR100 | average accuracy: 87.1 -> 88.5 | Improved |
| add usage penalty term to PromptFusion-Lite (ablation) | PromptFusion-Lite · Split-CIFAR100 | GFLOPs: 42.6 -> 38.3 GFLOPs; Acc 88.2 -> 87.3 | Improved |
| add knowledge distillation on the decision module F (ablation) | PromptFusion-Lite · Split-CIFAR100 | average accuracy: 87.0 -> 87.3 | Improved |
| use PromptFusion with Stabilizer (CoOp) and Booster (VPT) instead of prompt-pool methods | L2P, DualPrompt, CODA-Prompt · Split-ImageNet-R memory-free | average accuracy: 78.7, +5.3 over CODA-Prompt | Improved |
| use CoOp (Stabilizer) in place of VPT for retaining old tasks | prompt tuning for continual learning · Split-CIFAR100 pilot | average performance drop on old tasks: VPT 23.0% -> CoOp 9.9% | Improved |
| train a new VPT prompt set per task and concatenate (ablation) | Booster module · Split-CIFAR100 | average accuracy | Hurt |
| swap ViT backbone weights for CLIP visual encoder in VPT (ablation) | VPT · Split-CIFAR100 pilot | average accuracy | Hurt |
| use PromptFusion-Lite instead of PromptFusion | PromptFusion · Split-ImageNet-R | accuracy and GFLOPs: Acc 82.2 -> 81.5; GFLOPs 42.6 -> 34.4 | Mixed |
| vary memory buffer size from 1K to 5K (ablation) | PromptFusion · Split-CIFAR100, Split-ImageNet-R, Core50 | average accuracy | No clear effect |
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.
- DualPrompt — DualPrompt: Complementary Prompting for Rehearsal-Free Continual Learning (ECCV 2022) · better on 4 of 4 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
- Advancing Prompt-Based Methods for Replay-Independent General Continual Learning (ICLR 2025 · similarity 0.90)
- Mixture of Experts Meets Prompt-Based Continual Learning (NeurIPS 2024 · similarity 0.90)
- Evolving Parameterized Prompt Memory for Continual Learning (AAAI 2024 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.89)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.88)
- Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning (ICLR 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.
