World Editing: Intervening on Executable Worlds at Increasing Depth

World Editing. Intervening on Executable Worlds at Increasing Depth. By Max Ku and seventeen others, from the University of Waterloo, G-G-G, and two other institutions. Published on the first of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Interactive world models are getting better at generating environments and at acting inside them. Deliberately editing an existing executable world is a different matter, and it remains underexplored. The authors define world editing as intervening on an existing world while preserving whatever should stay unchanged. That second half is what separates editing from building something new. A change that works but breaks the rest of the world has not done the job.

The axis they introduce is intervention depth: how strongly an edit couples a world's entities, its dynamics and its systems. Tasks are spread along that axis, spanning interventions on properties, on entities, on dynamics, and on whole systems. Depth is meant as a measure of entanglement rather than of effort. The deeper an edit, the more of the world's parts it ties together, and the more there is that could go wrong around it.

The capability is instantiated through industry-grade game modding. Alongside IGMWorld comes IGMBench, a benchmark of one hundred ten tasks with more than eleven hundred executable criteria about state and behaviour, across Minecraft and Terraria. Edits are evaluated through deterministic checks of four kinds: whether the result executes, whether it behaves as asked, whether what should be preserved was preserved, and whether it looks right. Because the worlds are executable, each of these can be checked by running them, and the scores come from that rather than from a judgement call.

Frontier coding agents already show substantial world-editing capability. The strongest configuration solves seventy-eight point two percent of tasks under a strict task-level criterion, and at the level of individual criteria its performance reaches ninety-four point eight percent. Reliability generally falls as intervention depth rises. That pattern holds even among tasks with similar numbers of evaluation criteria, so it is not simply that deeper tasks carry more checks.

Most failed edits still build and load successfully. The main difficulty, the authors suggest, is making the edited world behave as requested. Getting the change to compile is the easy part. Getting the world to do the new thing, while still doing all the old things it should, is where agents fall short. Visual consistency is a separate weakness: every evaluated configuration stays below fifty percent joint visual pass rate.

Two conclusions follow from the results. World editing is a capability distinct from generating a world or interacting with one, and executable games give a practical testbed for studying it. Within that testbed, depth rather than the count of criteria is what tracks the edits agents get wrong.