Tunable MAGMAX Enables Preference-Aware Model Merging for Continual Learning
A novel framework named Tunable MAGMAX enables preference-sensitive management of task-specific performance in continual learning (CL). CL involves sequentially training models on various tasks while preventing catastrophic forgetting. Although recent developments utilize large pre-trained models (LPMs) and merging strategies like MAGMAX, current approaches tend to prioritize average performance, overlooking deployment contexts and user preferences. Tunable MAGMAX offers a preference vector that regulates how many elements are chosen from each task vector during the merging process, tailoring performance to specific deployment requirements. Additionally, the paper introduces an automated technique for selecting the preference vector. This research is available on arXiv (2605.20803).
Key facts
- Tunable MAGMAX is a model merging framework for continual learning.
- It enables preference-aware control of task-specific performance.
- A preference vector controls element selection from each task vector.
- The method adjusts merged model performance for different deployment needs.
- An automatic method for preference vector selection is proposed.
- The paper is published on arXiv with ID 2605.20803.
- Continual learning aims to train models sequentially on multiple tasks.
- Existing methods do not address varying user preferences.
Entities
Institutions
- arXiv