Fudan and Tongyi introduce ToolCUA for GUI-Tool path selection in agents
别光给Agent加Tool了,它根本选不明白!复旦×通义提出全新CUA训练范式
Fudan University and Tongyi Lab introduced ToolCUA-8B, which reaches 46.85% accuracy on OSWorld-MCP after training with about 4k synthetic tools and 180k interleaved GUI-Tool trajectory steps.
Why it matters: Fudan and the Qwen team train an agent to pick the right on-screen tool with 4,000 synthetic tools and 180,000 cross-trajectory steps, hitting 46.85% on OSWorld-MCP with an 8B model.