Reproduce “Detecting Silent Action-Interface Mismatch with Calibrated Relational World 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
Rui Zhu, Xurui Li, Haixu Tang
The authors detect semantic robot-command mismatches in offline replay by comparing state-action-next-state transitions with clean-learned relationships, then ranking inverse transformations. Across ten LIBERO Spatial tasks, at a matched 2.5% clean false-positive rate, ridge, joint GMM, and its conditional score flag 46.08%, 60.96%, and 65.66% of mutated rows. Joint scoring localizes better, but replay establishes neither safe correction nor deployment calibration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.