Reproduce “SAFE-Merge: Data-Free Continual Model Merging with General Knowledge Preservation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zihuan Qiu, Zhiyang Liao, Chiyuan He, Yi Xu, Fanman Meng, Linfeng Xu, Qingbo Wu, Hongliang Li
Researchers from the University of Electronic Science and Technology of China developed SAFE-Merge, a data-free continual model merging method that explicitly protects pretrained general knowledge while learning new tasks. It consistently achieved the highest H-score, balancing task accuracy and held-out generalization, improving over strong baselines by up to 8.1 points in H-score on vision tasks and 0.6 points on NLP tasks, without increasing inference costs.
Key insights
- Parameter singular directions can serve as effective data-free proxies for distinguishing general knowledge from task-specific information, enabling a quantifiable risk assessment for parameter updates.
- A balanced approach to continual merging requires explicitly safeguarding the foundation model's general knowledge, not just previously learned tasks, which can be achieved through risk-aware sparse masking.
- Masked low-rank recovery, constrained to 'safe' parameter regions, can effectively restore critical task information and plasticity without reintroducing interference or violating general knowledge preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.