WebVoyager
A public benchmark for browser agents that evaluates end-to-end task success rates on tasks based on real websites.
Timeline (last 90 days)
Comparing BF16 and NVFP4 under the same harness, WebVoyager dropped from 77.72 to 66.31, DeepSWE from 24.65 to 18.2, and MMLU-ProX from 76.91 to 73.76, with other rows within the noise range.
Evaluating browser agents typically uses public benchmarks such as WebArena, WebVoyager, Mind2Web, and GAIA, measuring end-to-end success rate and step efficiency on multi-step tasks on real websites.
Since 2024, as multimodal large model capabilities have improved, projects such as Browser Use, WebVoyager, and Anthropic's Computer Use have emerged in succession.
Some mature visual agent frameworks (such as SeeAct and WebVoyager) incorporate Set-of-Mark (SoM) technology, which annotates interactive elements on screenshots with numbered markers, converting coordinate-based positioning into a multiple-choice problem.
Representative browser Agent projects include WebVoyager, Agent-E, and commercial products Browserbase and Multion.
VLM 视觉方案（如 WebVoyager、SeeAct）存在幻觉问题、坐标定位精度不足，且每次操作需消耗大量 token 处理图像信息
视觉语言模型方案的代表性项目包括 WebVoyager 和 SeeAct，存在幻觉和坐标定位精度不足问题
All Facts (7)
Comparing BF16 and NVFP4 under the same harness, WebVoyager dropped from 77.72 to 66.31, DeepSWE from 24.65 to 18.2, and MMLU-ProX from 76.91 to 73.76, with other rows within the noise range.
50%UnverifiedEvaluating browser agents typically uses public benchmarks such as WebArena, WebVoyager, Mind2Web, and GAIA, measuring end-to-end success rate and step efficiency on multi-step tasks on real websites.
50%UnverifiedSome mature visual agent frameworks (such as SeeAct and WebVoyager) incorporate Set-of-Mark (SoM) technology, which annotates interactive elements on screenshots with numbered markers, converting coordinate-based positioning into a multiple-choice problem.
50%UnverifiedRepresentative browser Agent projects include WebVoyager, Agent-E, and commercial products Browserbase and Multion.
50%UnverifiedVLM 视觉方案（如 WebVoyager、SeeAct）存在幻觉问题、坐标定位精度不足，且每次操作需消耗大量 token 处理图像信息
50%Unverified视觉语言模型方案的代表性项目包括 WebVoyager 和 SeeAct，存在幻觉和坐标定位精度不足问题
50%UnverifiedSince 2024, as multimodal large model capabilities have improved, projects such as Browser Use, WebVoyager, and Anthropic's Computer Use have emerged in succession.
50%