Reproduce “When World Models Lie: Adaptive Safety Analysis Under Wrong Imaginations” 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
John Cao, Somil Bansal
The authors introduce a latent safety filter that compares a learned world model’s predicted state with the state inferred from the next observation. It widens plausible states from this mismatch and evaluates safety pessimistically. In open-loop Franka Panda replays with stale observations, it rescued 16 of 20 failures, versus 8 of 20 for LatentSafe; this does not guarantee live-deployment safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.