WorldBench

[Paper] [GitHub]

WorldBench is an execution-based benchmark and agentic judge designed to evaluate large language models on open-ended, LLM-generated Three.js voxel worlds.

The main repository, where WorldBench is actively developed, is on GitHub: KrishBakshi/worldbench. This Hugging Face repository is a release snapshot of the benchmark's data and evaluation logic.

Overview

Large language models can write complete, interactive 3D worlds as code, but traditional automated grading is unreliable. Existing judges typically rely on a single view: either a vision-language model scoring rendered snapshots or a language model reviewing the source code.

WorldBench introduces an interactive, dual-channel agentic judge that explores the running world in real-time controlling clocks, orbiting structures, and deploying navigator agents to frame each biome. while cross-referencing visual claims against actual source code execution.

Repository Contents

This dataset repository hosts the raw artifacts for the benchmark. Files are uploaded as-is; there is no load_dataset loader.

Path	Contents
prompts/prompt.md	The world prompt specification every model receives
tests/WC00N_*/	Rubrics and per-test definitions (test.yaml, README, scoring code) for the five tests, 280 points in total
eval/judge.yaml	Judge configuration: which model fills each judge role, rate limits, retries
eval/, checks/, harness/	Evaluation logic: structural gate, instrumented browser capture, LLM/VLM judges, scoring. harness/ also holds the generation loop
results/<model>/validation.json	Per-model scores with the judges' verdicts and quotes
assets/	Overview video

Running the evaluation

Put a model's output at inputs/<model>/world.html, copy .env.example to .env with your API keys, then:

uv run python -m eval.run <model>                      # the full ladder
uv run python -m eval.run <model> --test WC001,WC002   # a subset

Results are written to outputs/<model>/validation.json. Capturing needs Node (it drives a headless browser).

Citation

If you use WorldBench in your research, please cite our paper:

@misc{bakshi2026worldbench,
      title={WorldBench: Evaluating LLMs on Three.js Voxel World Generation}, 
      author={Krish Bakshi},
      year={2026},
      eprint={2610.10622},
      archivePrefix={arXiv},
      primaryClass={cs.GR}
}