Realtime-Bench: Benchmarking LLMs' Real-Time Interaction Capabilities
Understand the story
On November 2, 2026, Import AI issue 479 reported that researchers launched Realtime-Bench, a benchmark for evaluating large language models on real-time interactive tasks. The benchmark covers 50 gaming and web tasks, testing models on three capabilities: temporal state awareness, adaptive reasoning, and time-aware action selection. Results showed that GPT-6 Astra (high) led with a 36% pass rate under wall clock (real-time) settings; when the clock was paused, high-reasoning models achieved pass rates of 70%-80%. This gap reveals the current shortcomings of large models in real low-latency environments. The same issue also mentioned a US Defense Intelligence Agency space security report and Genie 4, among other items. No follow-up developments on this event have been reported so far.
AI overview of reporting · Updated 2 h前
Reporting timeline
Explore different aspects of the story through its reporting.
- Import AIImport AI 479: Spacewar; low-latency tasks versus LLMs; Genie 4
Researchers launched Realtime-Bench to evaluate large language models on real-time interactive tasks, covering 50 gaming and web tasks and testing three capabilities: temporal state awareness, adaptive reasoning, and time-aware action selection. In testing, GPT-6 Astra (high) topped the wall clock setting with a 36% pass rate, while high-reasoning models reached 70%-80% once the clock was paused. The same issue also covers a US Defense Intelligence Agency space security report and Genie 4.
Story heat trend
Not enough continuous observations to show a trend yet.