Research map / Continual learning / Domain and prompt-based
SAME: Sparse and Anchored Model Editing for Heterogeneous Incremental Learning under Limited Data
Zixuan Duan, Zeyu Zhang, Fengyuan Lu, Shaofeng Zhang, Wenbin Li, Qi Fan, Yang Gao
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 new knowledge injection (closed-form key-value fit) (ablation) | continual fine-tune · MTIL HIL Split-2 5-shot | average accuracy: 46.0 -> 59.9 | Improved |
| add parameter anchor (ablation) | SAME (NKI) · MTIL HIL Split-2 5-shot | average accuracy: 59.9 -> 65.3 | Improved |
| add knowledge anchor (ablation) | SAME (NKI+PA) · MTIL HIL Split-2 5-shot | average accuracy: 65.3 -> 65.5 | Improved |
| add sparse fine-tuning (ablation) | SAME (NKI+PA+KA) · MTIL HIL Split-2 5-shot | average accuracy: 65.5 -> 66.0 | Improved |
| use Fisher-based mask instead of random mask (ablation) | SAME sparse FT · MTIL HIL Split-2 5-shot | average accuracy: 66.0 -> 65.3 | Hurt |
| raise sparsity ratio from 30% to 40% (ablation) | SAME · MTIL HIL Split-2 5-shot | average accuracy: 66.0 -> 60.9 | Hurt |
| lower sparsity ratio from 30% to 20% (ablation) | SAME · MTIL HIL Split-2 5-shot | average accuracy: 66.0 -> 65.6 | Hurt |
| change lambda_p from 300 to 200 or 400 (ablation) | SAME · MTIL HIL Split-2 5-shot | average accuracy: 66.0 -> 65.8 | No clear effect |
| SAME vs best model fusion baseline (TIES-Merging) | SAME · MTIL HIL Split-2 5-shot | average accuracy: 61.7 -> 66.0 | Improved |
| SAME vs best continual learning baseline (MoE-Adapter) | SAME · MTIL HIL Split-2 5-shot | average accuracy: 59.2 -> 66.0 | Improved |
| SAME vs best continual learning baseline in Split-3 (MoE-Adapter) | SAME · MTIL HIL Split-3 5-shot | average accuracy: 56.9 -> 61.4 | Improved |
Closest papers by meaning
- Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning (NeurIPS 2026 · similarity 0.88)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.87)
- In-Context Editing: Learning Knowledge from Self-Induced Distributions (ICLR 2025 · similarity 0.87)
- Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Efficient Model Editing with Task-Localized Sparse Fine-tuning (ICLR 2025 · similarity 0.87)
- Dynamic Integration of Task-Specific Adapters for Class Incremental Learning (CVPR 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.
