Cram

New methods enhance LLM continual learning for multimodal tasks

Researchers have developed new methods to improve multimodal continual instruction tuning for large language models. CRAM focuses on isolating task-specific patterns into independent modules and using adaptive-rank inst…