Research map / Continual learning / Continual fine-tuning of LLMs
Merge to Remember: Sharpness-Aware Isotropic Merging for Continual Learning
Qun Yang, Enneng Yang, Wei Chen, Li Shen, Long Lan
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 |
|---|---|---|---|
| use sharpness-aware perturbation instead of random noise in SA-BCD (ablation) | SA-BCD · ViT-B/32 | average accuracy (%): 73.20 -> 82.74 | Improved |
| add sharpness-aware perturbation (vs none) (ablation) | SA-BCD · ViT-B/32 | average accuracy (%): 77.71 -> 82.74 | Improved |
| select parameters by momentum instead of instantaneous gradient (ablation) | SA-BCD · ViT-B/32 | average accuracy (%): 81.02 -> 82.74 | Improved |
| add block selection of top 30% parameters (vs update all) (ablation) | SA-BCD · ViT-B/32 | average accuracy (%): 79.54 -> 82.74 | Improved |
| use SA-BCD instead of AdamW fine-tuning (ablation) | SAIM · ViT-B/32 20 tasks | loss drift after 8 tasks: 1.4522 -> 0.0407 | Improved |
| add SA-BCD fine-tuning to TIES-Merging (ablation) | TIES-Merging · ViT-B/32 20 tasks | accuracy (%): 59.9 -> 63.4 | Improved |
| add SA-BCD fine-tuning to Task Arithmetic (ablation) | Task Arithmetic · ViT-B/32 20 tasks | accuracy (%): 60.0 -> 61.2 | Improved |
| add SA-BCD fine-tuning to simple averaging (SWA) (ablation) | SWA · ViT-B/32 20 tasks | accuracy (%): 61.1 -> 59.6 | Hurt |
| add SA-BCD fine-tuning to SAIM merging (ablation) | SAIM · ViT-B/32 20 tasks | accuracy (%): 70.8 -> 73.6 | Improved |
| use SAIM instead of EWC/ER at few tasks | SAIM · ViT-B/32 8 tasks | average accuracy (%): 84.7 (EWC) -> 82.1 | Hurt |
| use SAIM instead of OPCM | SAIM · ViT-L/14 20 tasks | average accuracy (%): 76.0 -> 87.5 | Improved |
| use SAIM instead of best PEFT method (SD-LoRA) | SAIM · ViT-B/32 20 tasks | average accuracy (%): 66.4 -> 73.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SD-LoRA — SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025) · better on 1 of 1 shared result
- InfLoRA — InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning (CVPR 2024) · better on 1 of 1 shared result
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
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.90)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.90)
- BECAME: Bayesian Continual Learning with Adaptive Model Merging (ICML 2025 · similarity 0.89)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.89)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.89)
- MINGLE: Mixture of Null-Space Gated Low-Rank Experts for Test-Time Continual Model Merging (NeurIPS 2025 · 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.
