ROSE targets lifelong knowledge editing in multimodal models
A Northwestern Polytechnical University research team proposed ROSE, a framework for continual knowledge editing in multimodal large language models. The team says ROSE constrains parameter-update directions and selectively invokes edited knowledge to reduce interference and forgetting. Its paper was selected as a NeurIPS 2026 Spotlight. In reported VLKEB tests using LLaVA-v1.5-7B, ROSE retained 95.24% reliability after 1,000 continuous edits, compared with 68.14% for FT-L.
Continual knowledge updates can cause multimodal models to forget earlier edits and alter answers to unrelated inputs. The report identifies these as central problems for lifelong knowledge editing, and the team’s reported ROSE results address both reliability and locality. The benchmark figures come from the research team and were not independently verified in the supplied report.
机器之心 2026-11-21 12:11 四川
在新知识学习与原有能力保持之间实现更好的平衡
现实世界中的事实知识并非一成不变:企业 CEO 可能更替,产品信息会随着发布与迭代而更新,赛事结果也会不断产生。多模态大模型需要及时吸收这些新信息,才能保持知识的时效性。然而,在持续引入新事实的过程中,如何避免新旧知识之间的相互干扰,并尽可能保留模型在预训练阶段形成的通用能力,仍是亟待解决的关键问题。