Deep Search in Real-World Web Environments

One-line takeaway

OpenDS-World evaluates deep-search agents on the live open Web: 200 human-verified tasks with unique answers spanning modalities, file formats and interactive access. Among 12 frontier models in one tool environment the best reaches 39.0% accuracy, versus 21.1% for humans over 1,000 attempts; tool coordination and long-horizon search remain hard.

OpenDS-World 在真实开放网络上评测深度搜索 agent：200 道经人工核验、答案唯一的难题，跨多种模态、文件格式和需要交互才能访问的页面。在统一工具环境下测了 12 个前沿模型，最好的准确率 39.0%；人类 1000 次独立作答为 21.1%。分析显示各模型的搜索策略差异很大，工具协调和长程搜索仍是难点。