Realtime-Bench Tests 12 Frontier Models on Dynamic Tasks
AIbase AI新闻 reports that NVIDIA and UT Austin introduced Realtime-Bench, a benchmark with 50 tasks across eight scenario types for assessing real-time perception, time-adaptive reasoning, and action timing. Researchers tested 12 frontier models under three environment-advancement clock protocols, and reportedly found that raising reasoning intensity did not reliably improve success. The report says GPT-6-Astra achieved a 44% task success rate when environments advanced according to actual reasoning time.
Realtime-Bench offers researchers a focused way to evaluate whether models can balance low latency, adaptive reasoning, and well-timed actions in dynamic environments.
大模型实时交互遭遇滑铁卢！英伟达等机构抛出 Realtime-Bench：最强 GPT-6-Astra 成功率仅 44%
AIbase基地
发布于AI新闻资讯 · 1 分钟阅读 · Nov 5, 2026
7
人工智能大模型在文本创作、代码编写和图像生成等静态任务上已经展现出惊人的实力，但在瞬息万变的实时交互场景中，它们的真实表现究竟如何?近日，英伟达携手得克萨斯大学奥斯汀分校（UT-Austin）等顶尖机构正式发布了全新的实时交互任务基准——Realtime-Bench。该基准的诞生，旨在为评估人工智能在动态环境下的综合表现提供一把精准的量尺。
#NVIDIA#University of Texas at Austin#Realtime-Bench#GPT-6-Astra