Sources
PAWBench finds no video world model reproduces the distribution of possible outcomes, only one plausible trajectory
The paper (arXiv 2608.27345, submitted 2026-08-27) argues that a world model should sample from the probability distribution of physical behaviors given the same initial observation and action, not just emit one plausible rollout. PAWEval evaluates eleven systems over 50 scenarios and concludes no model consistently matches reference probabilities while covering the range of valid behaviors. The authors tested prompt phrasing, noise sampling and training modifications as fixes, and none closed the gap, which sets a floor for anyone planning to use video generators as simulators for agent training.
↳ Follow the thread