StressDream: Steering Video World Models
for Robust Policy Evaluation and Improvement

Junwon Seo1, Sushant Veer2, Ran Tian2, Wenhao Ding2, Apoorva Sharma2, Karen Leung2,3, Edward Schmerling2, Marco Pavone2,4, Andrea Bajcsy1
1 Carnegie Mellon University   2 NVIDIA Research   3 University of Washington   4 Stanford University

We present StressDream, an inference-time method that optimizes diffusion noise to steer video world models toward plausible, high-impact outcomes — enabling robust evaluation and improvement of robotic policies.