Researchers introduce GraphForge, an evidence-graph based framework for synthesizing agent training data that grounds both tasks and their verification rubrics in real files rather than model-generated content. Starting from occupation-grounded seeds, the pipeline assembles a workspace of real files, builds an evidence graph over their relations, and uses a revision agent to repair tasks and rubrics before collecting trajectories. Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories improved GDPVal by 65.7 points and lifted scores on Workspace-Bench-Lite and SpreadsheetBench II by 7.7 and 13.7 points respectively, with further gains from rejection fine-tuning on the model’s own rollouts.
