Continual learning work targets forgetting in AI models

Researchers released papers on continual learning, including CP-MoE, a Mixture-of-Experts framework for large language models and vision-language models that aims to reduce catastrophic forgetting. CP-MoE uses a transient expert to capture early task-specific updates, guide routing toward compatible stable experts, and protect important historical parameters during merging. A separate survey argues that continual learning is shifting from parameter-centric model adaptation toward system-level adaptation across training, post-training, inference, and external components such as memory and skill libraries.