OpenAI 模拟部署降低评估意识至接近真实流量水平
OpenAI 分享了模拟部署的技术细节，能大幅降低评估误差，对研究模型评估和智能体部署的人很有参考价值。
OpenAI 在推文中介绍其模拟部署方法可将评估意识降低至接近真实生产流量水平。该方法扩展至带状态工具的智能体部署，表明工具模拟器在给定充分上下文和能力时能够生成逼真的轨迹。
这有助于更准确地评估模型在真实场景中的表现。
Simulated deployments also reduced evaluation awareness to levels close to real production traffic. We extended the method to agentic deployments with stateful tools, showing that tool simulators can produce realistic trajectories when given sufficient context and capabilities.📊 7 ⚡
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Key entities and technical topics under continuous tracking include OpenAI, 模拟部署, 智能体, 评估意识.
Source attribution: curated from OpenAI (@OpenAI). Cross-source corroboration: 11 independent sources have covered this development (including arXiv: OpenAI, berryxia, 向阳乔木), confirming industry-wide signal strength.
