As interactive 3D worlds become a common setting for studying intelligent behavior, researchers face a practical problem: finding anomalies in these simulated environments, such as floating objects, traversable walls, or objects that clash with their surroundings. A new paper introduces WorldAuditBench, a benchmark designed to test whether multimodal AI systems can automate this kind of auditing.
The authors argue that 3D world auditing demands the tight coupling of two capabilities: action, to navigate the world and search for anomalies systematically, and visual reasoning, to interpret multimodal observations and flag potential problems. Whether multimodal agents can actually combine these abilities — using reasoning to spot anomalies and actions to validate them — remains largely unexplored.
WorldAuditBench comprises 213 anomaly tasks across 13 environments built with Unreal Engine 5 and Three.js, covering five anomaly families. The team evaluated five frontier models under a fixed exploration budget using two paradigms: VLA-based exploration followed by VLM-based anomaly identification, and an end-to-end VLM agent in which visual reasoning directly guides action selection.
Across the evaluated models and both paradigms, success rates ranged from 6.6% to 42.3%, well below human performance of 83.4%. The authors present the benchmark as a testbed for studying how multimodal agents couple action and visual reasoning, while highlighting current limitations in gathering and interpreting evidence during exploration.