Towards Safe World Models: Risks, Defenses, and Evaluation
A survey maps world-model safety risks, defenses, and evaluation across six lifecycle stages using 163 works.
This survey argues world-model safety cannot be judged from predictive accuracy because errors in observation, memory, dynamics, planning, control, or feedback can drive unsafe decisions. It organizes 163 cited works around a six-stage lifecycle and separates non-adversarial failures, attacks, defenses, evaluation protocols, benchmarks, and deployment conditions. The framework highlights supported claims, unresolved attribution, and measurements still needed to link model behavior to closed-loop safety.
- Safety is organized across six lifecycle stages from data to feedback.
- The survey covers failures, attacks, defenses, protocols, and benchmarks.
- It flags gaps between model behavior and closed-loop operational safety.
From the original article
1 min read at arxiv.orgWorld models support planning, control, learning from imagined experience, and model-mediated evaluation by predicting action-conditioned consequences. Their safety cannot be inferred from predictive accuracy alone: errors in observations, state and memory, dynamics, evaluation, execution, or feedback can propagate into unsafe decisions and persist through updates.
By Xiaoma Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.