Continual Learning for VLMs

This field studies how vision-language models such as CLIP and multimodal large language models can learn new tasks, domains, or modalities while retaining earlier knowledge and cross-modal alignment. Common approaches include replaying examples or compact memories, regularizing parameter and representation changes, and using parameter-efficient prompts, adapters, low-rank updates, or routed experts. Researchers evaluate these methods in settings such as class- and domain-incremental recognition, visual question answering, multimodal instruction tuning, retrieval, and video understanding. A central tension is balancing learning new distributions against forgetting and preserving generalization, while benchmarks increasingly test task order, realistic distribution shifts, transfer, memory, and computational cost.