Reproduce “ROSE: Risk-Aware Orthogonal Subspace Navigation for Lifelong Knowledge Editing in Multimodal Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Lingyun Song, Ziyao Chen, Kang Pan, Xiaolin Han, Yifei Zhang, Xiaoqi Wang, Yudai Pan, Xuequn Shang

ROSE updates multimodal assistants while aiming to preserve earlier edits and avoid changing unrelated answers. During editing, it steers updates away from directions used by prior edits and limits where they act; at inference, it uses edited information only when inputs appear relevant. At 1,000 edits, it has higher Reliability than LiveEdit across three tasks, but lower multimodal Locality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.