GUI world models (GUI-WMs) predict future screens from the current screen and an action. The paper identifies state aliasing: identical screens can correspond to different underlying states, so observation-only predictions are unreliable. To diagnose this, the authors introduce StateAliasBench, a that isolates such ambiguities. They then propose lightweight predictive-state recovery that infers structured state from past history and feeds it into frozen GUI-WMs through a deterministic interface. Family-specific specialists handle different state types, and multi-teacher distillation combines them into a single estimator. Experiments show that existing GUI-WMs systematically fail under observation-only conditioning, while the proposed augmentation restores state-sensitive prediction, preserves generative fidelity, and improves downstream GUI agent performance on AndroidWorld. The work highlights that reliable world modeling must account for hidden transition state, not just visible observations.
GUI world models fail when hidden state is omitted from predictions
Authors diagnose state aliasing and propose predictive-state recovery to restore accuracy.
University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences · Huawei Noah's Ark Lab
In one line
State aliasing in GUI world models is resolved by inferring hidden state from history.
Before you read this
Three ideas this paper builds onThe authors identified a failure mode they call "state aliasing" in GUI World Models (GUI-WMs): because GUI-WMs typically condition only on the current screenshot and action, they cannot distinguish between situations where the visible interface looks the same but the underlying environment state (e.g., clipboard contents, active timer) differs.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.