Skip to content
Xinzhiyuan · WeChat

USTC Papers Study Lifelong Learning for LMMs via Multimodal Knowledge Injection

教大模型终身学习!中科大连发两篇顶会,突破「知识注入」双重困境

USTC researchers released MMEVOKE and KORE: MMEVOKE contains 9,422 samples across 159 subcategories, while KORE uses knowledge-tree augmentation and null-space constrained fine-tuning to reduce catastrophic forgetting during multimodal knowledge injection.

Why it matters: A USTC team's MMEVOKE benchmark and KORE fine-tuning method offer reusable evaluation and training ideas for forgetting in continual learning of multimodal models.

Read the original ↗Export Markdown