Beyond Next-Screen Prediction: The Decision Value of GUI World Models

Key points

• We study this attribution problem through controlled interventions on GUI action selection.1
• 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.1
• 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.1
• 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.1

Sources (1)

• [1]Beyond Next-Screen Prediction: The Decision Value of GUI World Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 02:03 AM

  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.

Extractive summary: sentences quoted from the sources.