Interactive world models can generate environments and act within them, yet purposeful editing of existing executable worlds remains underexplored. We define world editing as intervening on a given world while preserving properties that must stay unchanged, and introduce the notion of intervention depth to describe how strongly an edit couples entities, dynamics, and systems. By leveraging industry‑grade game modding we instantiate this ability in IGMWorld and release IGMBench, a benchmark comprising 110 tasks and over 1.1K executable state and behavioral criteria across Minecraft and Terraria. The tasks cover property, entity, dynamics, and system interventions and are evaluated on deterministic executability, behavioral conformity, preservation, and visual consistency. State‑of‑the‑art coding agents already show substantial world‑editing capability: the strongest configuration solves 78.2% of tasks under a strict task‑level criterion and reaches 94.8% at the criterion level. Reliability generally drops with increasing intervention depth, a pattern that persists even among tasks with similar numbers of evaluation criteria. Most failed edits still build and load successfully, indicating that the main difficulty lies in making the edited world behave as requested. Visual consistency remains a separate weakness, with all configurations below a 50% joint visual pass rate. These findings demonstrate that world editing is a distinct capability from world generation and interaction, and that executable games provide a practical testbed for its study.
Review