Reproduce “ReferClaw: Disambiguating User Intent for Egocentric Assistance” 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
Hangyu Qin, Junbin Xiao, Angela Yao
The authors present ReferClaw, a video assistant that grounds candidate interpretations in video evidence and asks a targeted question when those interpretations imply different answers. It then answers using evidence tied to the user-selected interpretation. On a 6,240-question augmented BuddyVQA test set, frozen ReferClaw scores 66.2% accuracy versus 60.5% without asking; intent is benchmark-annotated, not verified user intent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.