2026-10-02 · America/Los_Angeles · 论文 · #12

GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis

Working agents need to read diverse files, coordinate tools, and produce deliverables. Training such agents requires tasks built on many real files with verifiable results, but few pipelines exist to synthesize this kind of data. Existing pipelines either generate files with models, which lack realism and diversity, or build tasks on real files without task-specific verifiers, leaving result quality unchecked. We introduce GraphForge, an evidence-graph based framework that grounds both the task and its verification in real files. Starting from occupation-grounded seeds for controlled…

热度 51.3 / 100；排名与评分保留该期记录。