Research map / Continual learning / Domain and prompt-based
Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation
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 KFF, a class-aware domain knowledge fission and fusion framework with prompt pools for continual test-time adaptation of ViTs, reporting state-of-the-art error on ImageNet-C and CIFAR-C benchmarks.
Problem
Continual test-time adaptation under irregularly switching, possibly repeating domains suffers from forgetting, insufficient learning of new knowledge, and interference from harmful historical domain knowledge, with unbounded growth of stored models or prompts.
Approach
Uses a paired domain prompt pool and class-aware prompt pool on a frozen ViT: knowledge fission creates new domain/class prompts when no existing prompt matches (via statistics distance and pseudo-label cosine similarity), and knowledge fusion greedily merges learned prompts into the pool (with MST-based clustering and nearest-pair merging) to bound pool size.
Key results
- On ImageNet-to-ImageNet-C (non-repeating, severity 5), KFF achieves 34.8% mean error versus 39.9% for DPCore, the previous best.
- Under 10 repeated rounds on ImageNet-C, KFF reaches 34.5% mean error, a 9.9% improvement over DPCore (44.4%).
- KFF uses about 0.09M learnable parameters, roughly 0.1% of the model, and in the repeating setting DPCore uses about five times more parameters in the final round.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KFF instead of DPCore | ImageNet-to-ImageNet-C non-repeating CTTA · ImageNet-to-ImageNet-C, non-repeating | error rate (%): 39.9 -> 34.8 | Helped |
| use KFF instead of CoLA | ImageNet-to-ImageNet-C repeating CTTA · ImageNet-to-ImageNet-C, 10 repeated rounds | mean error rate (%): 38.9 -> 34.5 | Helped |
| use KFF instead of DPCore | CIFAR100-to-CIFAR100-C CTTA · CIFAR100-to-CIFAR100-C | mean error rate (%): 25.1 -> 22.5 | Helped |
| use KFF instead of C-MAE | CIFAR10-to-CIFAR10-C CTTA · CIFAR10-to-CIFAR10-C | mean error rate (%): 12.6 -> 12.4 | Helped |
| add domain prompts with KFI and KFU (ablation) | KFF without class prompts · ImageNet-to-ImageNet-C | mean error rate (%): 55.8 -> 39.5 | Helped |
| add class prompts with KFI and KFU (ablation) | KFF without domain prompts · ImageNet-to-ImageNet-C | mean error rate (%): 55.8 -> 50.9 | Helped |
| add KFI and KFU (ablation) | class and domain prompt model · ImageNet-to-ImageNet-C | mean error rate (%): 62.9 -> 34.8 | Helped |
| add KFU (ablation) | class and domain prompts with KFI · ImageNet-to-ImageNet-C | mean error rate (%): 36.9 -> 34.8 | Helped |
| use threshold-based selection instead of top-3 matching (ablation) | KFF class prompt selection · ImageNet-to-ImageNet-C | mean error rate (%): 35.1 -> 34.8 | Helped |
| use threshold-based selection instead of top-1 matching (ablation) | KFF class prompt selection · ImageNet-to-ImageNet-C | mean error rate (%): 35.4 -> 34.8 | Helped |
| use threshold-based selection instead of top-5 matching (ablation) | KFF class prompt selection · ImageNet-to-ImageNet-C | mean error rate (%): 35.9 -> 34.8 | Helped |
| vary domain and class prompt pool sizes within a reasonable range (ablation) | KFF · ImageNet-to-ImageNet-C | error rate | No clear effect |
Surprising findings
- Ablation shows that adding class and domain prompts without KFI and KFU degrades error to 62.9%, far worse than the source model's 55.8%.
Limitations
- Relies on access to source domain statistics during adaptation, which may be unavailable for privacy-restricted data.
- Evaluated only on synthetic, manually designed corruptions, not naturalistic shifts.
- Adds computational overhead during adaptation, which may be challenging for edge devices.
Experimental setup
Datasets: ImageNet-C CIFAR10-C CIFAR100-C
Benchmarks: ImageNet-to-ImageNet-C CIFAR100-to-CIFAR100-C CIFAR10-to-CIFAR10-C
Metrics: classification error rate (%)
Baselines: Tent SAR POEM CoTTA VDP RoTTA C-MAE ROID ViDA CoLA PALM DPCore
Builds on: DPCore CoTTA EATA VDP
Closest papers by meaning
- BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation (ICML 2024 · similarity 0.89)
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.89)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.88)
- A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability (CVPR 2024 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.86)
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.