Deep research agents are increasingly required to retrieve precise information from massive corpora, yet multi‑hop, long‑horizon tasks remain hard. Synthetic data has become a scalable remedy, but existing pipelines simply expand seed tasks to increase apparent depth and breadth. The added evidence often couples weakly with the answer, causing a gap between intended and actual difficulty. GRADS addresses this by introducing a graph‑guided agentic data synthesis framework with explicit difficulty control. It runs a closed loop of seed selection, evidence expansion, question‑answer construction, and verification while keeping a graph that records the evidence and inference dependencies needed for each answer. At every step generation is conditioned on the relevant subgraph, preserving indispensable reasoning without accumulating irrelevant context. Experiments span security, software engineering, and scientific‑document domains. GRADS consistently produces verified tasks whose difficulty distributions are empirically separated when measured by independent solvers. Under equal training budgets, post‑training on this data yields gains at every stage: SFT raises the solution‑termination rate above 95% and cuts tool calls and reasoning turns by up to 75%, RL further reduces tool usage by about 19.5%. On held‑out, domain‑specific benchmarks the resulting 4B model matches Sonnet 4 on software engineering and surpasses it on scientific and security tasks.
Review