WorldCoder-Bench: Benchmark for Physically Grounded 3D Scene Generation

WorldCoder-Bench is a core task dataset for evaluating large language models on generating interactive, physically grounded 3D web scenes from natural language using Three.js. The benchmark, authored by shuolucs, was last updated on 2026-06-09. Generated programs must integrate 3D assets and obey spatial and physical constraints while keeping user controls synchronized with runtime state.

Use Cases

• Benchmarking LLM performance on generating interactive 3D scenes based on natural language descriptions.
• Evaluating model adherence to spatial and physical constraints in synthesized Three.js programs.
• Testing the synchronization of user-facing controls with hidden runtime state in generated interactive mechanics.

Strengths

• Focuses on physically grounded 3D world synthesis, a specific and challenging task for LLMs.
• Explicitly requires generated programs to integrate 3D assets and obey spatial constraints.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.

Provenance

Source

huggingface user shuolucs

Collection Method

Likely created as a benchmark task dataset for evaluating LLMs.

Freshness

Last updated 2026-06-09 03:21:24; freshness should be verified.