Hidden Forgetting in Continual Multimodal Learning: When Accuracy Survives but Grounding Fails

Researchers have identified a new failure mode in multimodal AI called hidden forgetting, where models retain the ability to answer questions correctly despite losing their ability to link those answers to visual input. While these systems appear to perform well on standard accuracy benchmarks, they lose the underlying grounding necessary to maintain reliable connections between text and imagery as they learn new tasks. This finding suggests that current evaluation methods may overestimate the stability of multimodal models during continuous training.