Deep research agents are increasingly used to pull precise information from large corpora, but they still stumble on tasks that require multiple reasoning steps over long horizons. According to a new paper on arXiv, synthetic data generation has become a promising way to create training examples at scale, yet existing approaches have limits.
The paper introduces GRADS, a method for graph-regulated agentic data synthesis with controllable difficulty. By using graph structure to guide the generation process, GRADS aims to produce training data that better matches the complexity of real-world multi-hop queries, while giving researchers a dial to adjust how hard the generated tasks are.
The abstract is brief and does not include experimental results, so the practical gains remain to be seen. Still, the framing points to a growing interest in making synthetic data not just plentiful, but also structurally meaningful for the long-horizon reasoning that trips up current agents.