Google and Artificial Analysis launch MKA agent benchmark
Google DeepMind and Artificial Analysis launched MKA, a benchmark for testing agents on visually complex, real-world knowledge work such as blueprints, charts, reports, and spreadsheets. The evaluation set contains 200 private tasks developed with experts across 10 industries and 83 occupations, while 10 sample tasks are public on Hugging Face. Google reported that Gemini 4 Carbon ranked second in quality, with an average task cost about one-sixth that of the leading model.
MKA adds an evaluation focused on professional work that requires interpreting visual files and producing multiple deliverables, and it is now available through Artificial Analysis for developers to test their own agents.
IT之家 11 月 11 日消息,当地时间 11 月 10 日,谷歌宣布联合 AI 模型评测机构 Artificial Analysis 推出多模态知识智能体基准测试 MKA(Multimodal Knowledge Agent),用于评估 AI 智能体处理视觉复杂文件、完成现实世界知识工作任务的能力。
谷歌表示,其最新发布的 Gemini 4 Carbon 在该测试中取得了参测模型第二名的成绩,完成任务的平均成本仅为排名第一模型的约六分之一。

谷歌 DeepMind 高级产品经理 Rohan Doshi 和首席工程师 Ben Caine 介绍称,现有不少知识工作基准测试仅依赖文本即可完成,而现实工作往往涉及大量视觉信息。例如,建筑师需要交叉核对建筑蓝图,工程师需要分析工程示意图,专业人员还需要将不同来源的信息整理成演示文稿、报告和电子表格。
#Google DeepMind#Artificial Analysis#MKA#Gemini 4 Carbon#Stirrup#Hugging Face