Ziyan Jiang, Jingbo Yang, Jiabao Ji, Yujian Liu, Qiucheng Wu, Tommi Jaakkola, Yang Zhang, Shiyu Chang
The paper introduces WorldAuditBench, a benchmark for 3D world auditing, and evaluates five frontier models under two auditing paradigms. The results show success rates ranging from 6.6% to 42.3%, highlighting current limitations in multimodal agents' ability to gather and interpret evidence. This work provides a testbed for studying action and visual reasoning in interactive 3D environments.
The development of efficient pipelines for identifying anomalies in simulated environments is crucial for studying intelligent behavior. WorldAuditBench has practical significance in applications such as game development, robotics, and virtual reality. Improving multimodal agents' ability to couple action and visual reasoning can lead to more realistic and engaging interactive 3D worlds.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.