New preprint: verifiable deterministic world models

Our exploration of world models for system assurance resulted in a semi-predictable but currently unfashionable choice: removing randomness and uncertainty from the latent space made world models more verifiable (although a tiny bit less picture-perfect). More surprisingly, this step made the behaviors produced by them more relevant to the real world.

As a result, we were able to transfer guarantees obtained on verifiable deterministic world models to the real image-based system. Intriguing!

Citation:

• Yuang Geng, Zhuoyang Zhou, Zhongzheng Zhang, Siyuan Pan, Hoang-Dung Tran, Ivan Ruchkin.
  Deterministic World Models for Verification of Closed-loop Vision-based Systems [Arxiv] [Github].
  Preprint, 2025.