OpenDS-World tests agents on real-web deep search
A multi-institution team introduced OpenDS, a deep-search task paradigm, and OpenDS-World, a 200-task benchmark for research agents operating across the open web. According to 机器之心, the tasks cover four modalities, 42 resource formats and seven evidence-access requirements; GPT-6 Astra led the 12 tested models with 39.0% accuracy, while Claude Opus 5.5 and Gemini 3.8 Flash scored 38.0% and 36.5%. The evaluation also records tool failures, execution time, search termination decisions and other execution traces.
OpenDS-World provides a more demanding test of evidence discovery and synthesis across multimodal, heterogeneous and interactive web resources, with results showing that failed tool calls and long execution paths consume substantial search budgets.
机器之心 2026-10-30 18:33 北京
新的深度搜索任务范式 OpenDS 及面向真实开放网络环境的评测基准 OpenDS-World
近年来，随着大语言模型推理和使用工具的能力不断提升，Deep Search Agent 已具备一定的自主搜索与信息整合能力。越来越多的研究开始尝试让 Agent 自主规划搜索路径，处理需要跨越多个信息源的复杂问题。
但当 Deep Search 真正进入现实互联网，研究过程会变得复杂得多。
#OpenDS#OpenDS-World#GPT-6 Astra#Claude Opus 5.5#Gemini 3.8 Flash#Qwen3.8 Max#GLM-5.3 Flash#DeepSeek V4.1-Flash#Kimi K3#Qwen3.8-Flash#GPT-5.6 Sol#Qwen3.8-27B