新闻
WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents
arXiv cs.AI · 发布于 · 阅读约3分钟
30秒读懂
- 发生了什么
- WorldAuditBench is a benchmark with 213 anomaly detection tasks across 13 Unreal Engine 5 environments for testing multimodal AI agents on 3D world auditing.
- 为何重要
- Relevant for engineers building or evaluating vision-language models and agents that must navigate and reason about interactive 3D environments systematically.
- 注意
- Current frontier models achieve only 6.6 to 42.3 percent success versus 83.4 percent human performance, indicating substantial gaps in coupling action and visual reasoning.
- agent
- language model
这条新闻背后的模式
每个模式都讲清楚技术如何运作、何时值得投入，以及在哪里会失效。
The Agent Architect
每周一个模式、一个权衡、一个生产事故案例。为构建智能体系统的人准备的每周简报。
每周一封邮件,一键退订。您的地址仅用于发送简报。