Skip to content
NewA calmer, faster ZeroHour

World Editing: Intervening on Executable Worlds at Increasing Depth

Hugging Face daily papersby Max Ku, Nok-Kan Law, Yu-Chien Tang ()1 min read3
AI summary · grok-4.7Summaries by Vani AI

Frontier coding agents edit Minecraft and Terraria worlds, but reliability falls as intervention depth grows.

The paper frames world editing as changing an executable world while preserving properties that should stay fixed, using intervention depth to describe how tightly an edit couples entities, dynamics, and systems. IGMBench contains 110 tasks and more than 1,100 executable criteria in Minecraft and Terraria. The strongest frontier coding-agent configuration solves 78.2% of tasks, with 94.8% criterion-level performance. Reliability declines as edits go deeper, most failures still build and load, and every configuration stays below a 50% joint visual pass rate.

Key facts
  • IGMBench covers 110 tasks and over 1,100 criteria in Minecraft and Terraria.
  • Best configuration solves 78.2% of tasks and 94.8% of criteria.
  • Success generally drops as intervention depth increases, even with similar criterion counts.
  • Failed edits usually still build and load; joint visual pass rates stay under 50%.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at huggingface.co

Interactive world models are increasingly capable of generating environments and acting within them, yet deliberately editing an existing executable world remains underexplored.

Hugging Face daily papers

By Max Ku, Nok-Kan Law, Yu-Chien Tang · huggingface.co

Read the full article at Hugging Face daily papers

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.