Reproduce “Tracking the risk of a deployed model and detecting harmful distribution shifts” 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
Aleksandr Podkopaev, Aaditya Ramdas
The authors propose monitoring model risk rather than input change, warning only when estimated target error exceeds source error by a chosen tolerance. The method compares a source-risk upper bound with an incoming-risk lower bound and uses confidence sequences to control false alarms during continuous monitoring. In corruption experiments, translation and severe fog worsened outcomes; labeled target data are required.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.