HyperAIHyperAI

Command Palette

Search for a command to run...

ExeWorldBench: Executable 3D World Modeling Benchmark Dataset

ExeWorldBench is a benchmark dataset for executable 3D world modeling released in 2026 by the University of Science and Technology of China in collaboration with The Chinese University of Hong Kong. The related paper is “ExeWorldBench: Benchmarking Code-Based Executable 3D World Modeling”. It aims to evaluate AI models’ understanding of mechanical structures, physical motion, and causal relationships in 3D worlds, as well as their ability to construct executable 3D scenes through code.

The dataset is built on the Blender 3D modeling and simulation environment and covers three types of tasks: mechanism reproduction, multi-stage mechanical system composition, and intent-driven executable world generation. The evaluation not only focuses on the visual structure of generated scenes but also examines the connections between mechanical components, motion processes, causal interactions, and goal completion, so as to measure models’ comprehensive ability to go from visual observations or text instructions to executable 3D world construction.

The dataset provides resources such as task instructions, reference images, motion reference materials, and evaluation prompts. It supports evaluation via Agent Harness or API, and can be used for research on evaluating code generation, physical reasoning, world models, and agents’ 3D scene construction capabilities.

This dataset centers on whether a runnable 3D world can be constructed, rather than merely focusing on reproducing the world’s appearance. It can be used to test a model’s ability to build executable worlds through code.

Dataset Composition

The dataset mainly contains the following directories:

t1: Mechanism Executability task, requiring the model to infer mechanical structure, component relationships, and motion from reference images, and to reconstruct an executable mechanism using Blender code;

  • t2: Compositional Executability task, requiring the model to build a system containing multiple mechanical stages, examining whether the spatial layout, mechanical connections, and causal interactions among components can form a complete execution process;
  • t3: Intent-Driven Executable World Synthesis task, requiring the model to build an executable 3D world based on high-level textual goals and physical constraints, covering two task forms: specified-component constraints and goal constraints;
  • api_prompt: provides API evaluation prompts corresponding to the three task types. T1 and T2 use a unified prompt template, while T3 provides separate prompts for different tasks;
  • eval: contains prompts used by evaluation models to assess the visual structure, mechanical motion, executable causal mechanisms, and goal completion of generated scenes.

Citation

@article{peng2026exeworldbench,
  title={ExeWorldBench: Benchmarking Code-Based Executable 3D World Modeling},
  author={Peng, Xinge and Fang, Zhen and Wang, Ruibing and Lv, Yongxuan and Li, Haodong and Zhang, Guohui and Ma, Xiaoxiao and Lu, Yiting and Chen, Lin and Che, Zehui and others},
  journal={arXiv preprint arXiv:2610.15166},
  year={2026}
}

Build AI with AI

From idea to launch — accelerate your AI development with free AI co-coding, out-of-the-box environment and best price of GPUs.

AI Co-coding
Ready-to-use GPUs
Best Pricing

HyperAI Newsletters

Subscribe to our latest updates
We will deliver the latest updates of the week to your inbox at nine o'clock every Monday morning
Powered by MailChimp