Researchers present Iris-mini and Iris-pro, two search agents trained at 35B and 397B parameter scales, along with the full data pipeline and training recipe used to build them. The training data is constructed by reverse-engineering multi-hop question chains from a web corpus’s hyperlink structure, ensuring questions cannot be solved by simple string matching, and the agents are refined through an alternating supervised fine-tuning and reinforcement learning procedure the authors call SFT-RL climbing. On benchmarks including BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE, the two models achieve the strongest reported results among open-source search agents at their respective parameter scales, using a single ReAct agent with no sub-agents or test-time verification.
