The paper proposes Agent-Editing World Model (AEWM), which models how reasoning and actions shape future task progress instead of predicting environment observations. AEWM uses an Action Judge to classify decisions as Critical, Exploratory, or Noisy, and State Revision to edit noisy ones. EditAct integrates these with real execution, directly changing the state underlying subsequent decisions rather than just critiquing.
AEWM is trained across Search, Terminal, and Software Engineering domains. On the Action Judge benchmark it reaches 70.5% macro-F1, 10.6 points better than the strongest frontier baseline. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2-6.7 points. Rejection sampling fine-tuning on verified trajectories, AEWM-RFT, further improves over Self-RFT by 2.2-2.6 points across three domains.