Research map / Continual learning / Domain and prompt-based
CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAPrompt, a class-incremental learning method that replaces task-ID prediction with a weighted aggregation of task-specific prompts, refined by a cyclic weight-prediction procedure and constrained by concave and linear losses.
Problem
Prompt-based class-incremental learning with pre-trained ViTs selects task-specific prompts via predicted task IDs; wrong task-ID predictions cause training/inference prompt inconsistency and forgetting.
Approach
Aggregate all task prompts with class-probability-derived task weights in both training and inference (no task-ID selection), justify lower error via Jensen's inequality under a concave condition enforced by concave and linear constraints, and iteratively refine the aggregation weights in cycles.
Key results
- CAPrompt reaches 95.52% ACC on CIFAR-100, 79.93% on ImageNet-R and 88.99% on CUB200 with ImageNet21K ViT-B/16, reported as 2-3% above the second-best method.
- Ablation on ImageNet-R: prompt aggregation raises ACC from 74.51% to 78.28%; cyclic weight adds 0.88%; concave and linear constraints add 0.61% and 0.16%.
- Adding more cycles (num>2) gives an additional 1.5%-2% improvement according to the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAPrompt instead of HiDe-Prompt | Pre-trained ViT-B/16 CIL · CIFAR-100 | ACC: 92.64 ± 0.22 -> 95.52 ± 0.12 | Helped |
| Use CAPrompt instead of ConvPrompt | Pre-trained ViT-B/16 CIL · ImageNet-R | ACC: 77.94 ± 0.06 -> 79.93 ± 0.19 | Helped |
| Use CAPrompt instead of HiDe-Prompt | Pre-trained ViT-B/16 CIL · CUB200 | ACC: 86.84 ± 0.14 -> 88.99 ± 0.15 | Helped |
| Use CAPrompt instead of HiDe-Prompt | Pre-trained ViT-B/16 CIL · ImageNet-R | AF: 2.88 ± 0.19 -> 3.37 ± 0.41 | Hurt |
| Add prompt aggregation (ablation) | ImageNet-R Base · ImageNet-R | ACC: 74.51 -> 78.28 | Helped |
| Add prompt aggregation (ablation) | ImageNet-R Base · ImageNet-R | AF: 3.01 -> 4.41 | Hurt |
| Add cyclic prompt weight prediction (ablation) | ImageNet-R aggregation model · ImageNet-R | ACC: 78.28 -> 79.16 | Helped |
| Add cyclic prompt weight prediction (ablation) | ImageNet-R aggregation model · ImageNet-R | AF: 4.41 -> 4.52 | Hurt |
| Add the concave constraint (ablation) | ImageNet-R aggregation and cyclic model · ImageNet-R | ACC: +0.61% | Helped |
| Add the concave constraint (ablation) | ImageNet-R aggregation and cyclic model · ImageNet-R | AF: 0.82% reduction | Helped |
| Add the linear constraint (ablation) | ImageNet-R model with aggregation, cyclic weighting, and concave constraint · ImageNet-R | ACC: +0.16% | Helped |
| Add the linear constraint (ablation) | ImageNet-R model with aggregation, cyclic weighting, and concave constraint · ImageNet-R | AF: 0.33% reduction | Helped |
Surprising findings
- Prompt aggregation increases average forgetting on ImageNet-R (3.01 to 4.41) even while accuracy rises, which the authors attribute to harder knowledge acquisition.
Limitations
- The concave condition cannot be exactly guaranteed by the constraints given the large parameter count, so the theorem's assumption is only approximately met.
- Average forgetting increases with the aggregation strategy on ImageNet-R.
- Additional cycles increase inference cost beyond two forward passes.
Experimental setup
Datasets: CIFAR-100 ImageNet-R CUB200
Benchmarks: CIFAR-100 ImageNet-R CUB200
Metrics: final accuracy (ACC) average forgetting (AF)
Baselines: L2P DualPrompt CODA-Prompt HiDe-Prompt EvoPrompt ConvPrompt CPrompt InfLoRA EASE SSIAT
Builds on: L2P DualPrompt CODA-Prompt HiDe-Prompt Prefix Tuning
Closest papers by meaning
- Consistent Prompting for Rehearsal-Free Continual Learning (CVPR 2024 · similarity 0.89)
- All You Need is One: Capsule Prompt Tuning with a Single Vector (NeurIPS 2025 · similarity 0.88)
- PEARL: Input-Agnostic Prompt Enhancement with Negative Feedback Regulation for Class-Incremental Learning (AAAI 2025 · similarity 0.88)
- Evolving Parameterized Prompt Memory for Continual Learning (AAAI 2024 · similarity 0.88)
- Training Consistent Mixture-of-Experts-Based Prompt Generator for Continual Learning (AAAI 2025 · similarity 0.87)
- C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning (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.