arxivcs.AI2026-07-02
Hidden Forgetting in Continual Multimodal Learning: When Accuracy Survives but Grounding Fails
Qianyu Chen, Canran Xiao, Runxuan Tang
Multimodal large language models must continually adapt to evolving tasks and domains, yet standard continual learning metrics mainly measure whether old answers remain correct, leaving the stability of multimodal grounding largely unexamined. We study this overlooked failure mod…