CaliBench: Are the Stochastic Dynamics of Video World Models Physically Calibrated?
The episode discusses 'CaliBench,' arguing that video world models often lack physically calibrated stochastic dynamics. Hosts analyze how models default to visually obvious features rather than reliably generating outcomes naturally. The discussion concludes that improving reliability requires incorporating explicit physical constraints and quantifying uncertainty into the training process.