arXiv · Computation and Language· Qisheng Su, Hanchen Wang, Guanru Zhu, Huicheng Jiang, Qiuyinzhe Zhang, Kou Shi, Zhen Fang, Ziao Zhang, Qingnan Ren, Zehui Chen, Tao Gui, Feng Zhao·· 3 小时前AI 评分58
GraphForge:基于图锚定工作空间的智能体训练框架
GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis
AI 导读
GraphForge 以证据图锚定任务与验证,基于真实文件组装工作空间并生成可执行任务与评分标准。Qwen3.6-27B 在 2,169 条 GraphForge 轨迹上微调后,GDPVal 达 1445.7(+65.7),Workspace-Bench-Lite 63.7(+7.7),SpreadsheetBench II 24.0(+13.7)。
来源:arXiv · Computation and Language · arxiv.org