Morning Edition · Tuesday, October 27, 2026Published at 2:18 AM EDT · New York

A new paper measures whether a pretrained world model reproduces the effect of small action changes

The authors compare matched simulator and model responses using Gram matrices at three perturbation scales, testing a property that video-prediction benchmarks do not measure.

A paper posted to arXiv on October 27 asks a narrow and useful question about learned world models: when you change an action slightly, does the model's predicted outcome change the way the real system does? The authors compare matched simu…