← All benchmarks

WorldAuditBench

Multimodal · 2026-09-30

213 interactive 3D world-auditing tasks in 13 Unreal Engine 5 and Three.js environments, each with one implanted anomaly from five families (static physics, interactive physics, spatial, temporal and semantic consistency). A multimodal agent explores the scene with first-person views under a 40-action budget, using MCP tools to move, look, interact, recall earlier frames and report, and receives an anomaly-type hint and an in-context example; a GPT-6 Astra judge checks the report and its cited evidence against the task rubric. The score is the percentage of tasks whose target anomaly is correctly identified.

Top models (higher is better)