Gui models improve action choices without extra screen predictions
Beyond Next-Screen Prediction: The Decision Value of GUI World Models
Artificial Intelligence
Summary
Picking the best action on a phone or computer screen often involves guessing what will happen next. This paper shows that directly scoring actions based on past examples can work better than predicting future screens first, which is what many current methods do. The authors created a new way called Direct Action Ranking (DAR) that avoids future screen predictions but still helps choose actions more accurately. They found that DAR performs better and even works well on apps it hasn't seen before, suggesting that focusing on scoring actions directly is more valuable than just predicting what happens next.
What this means in practice
- •For mobile app developers: Improve automated testing tools by directly ranking UI actions instead of relying on next screen predictions, increasing task success rates.
- •For software qa teams: Deploy smarter UI automation that better prioritizes actions across various apps without retraining, saving time in test design.
Authors
Sizhe Tang, Hanhan Zhou, Jeffrey Svajlenko, Rashmi Gangadharaiah
Abstract
GUI world models are increasingly used to improve action selection by predicting the consequences of candidate actions, but it is unclear how much of the resulting gain comes from prediction itself rather than from the evaluator that interprets those predictions. We study this attribution problem through controlled interventions on GUI action selection. With candidates and predictions held fixed, lookahead gains can change with how the prediction is presented to the evaluator, so end-to-end gains cannot be attributed to prediction alone. We then introduce Direct Action Ranking (DAR), which learns from reference actions to score candidates without generating future screens, and compare it with a prediction-augmented variant under matched supervision. Predicted successors add no detectable ranking signal once evaluation is learned directly. Without retraining, DAR improves AndroidWorld task success by 5.8--16.4 points across three proposers, outperforms world-model lookahead, and transfers zero-shot to unseen apps, where every prediction-conditioned configuration falls below proposer confidence. These results suggest that GUI world models should be evaluated by their decision value, the incremental action-selection value of their predictions beyond a strong direct evaluator, rather than by prediction fidelity or end-to-end lookahead gain alone.