We introduce Iris‑mini and Iris‑pro, two search agents trained at the 35B‑A3B and 397B‑A17B scales, and release the accompanying data pipeline and training recipe. Tasks are reverse‑engineered from the hyperlink structure of a web corpus: an entity graph is distilled from a seed page and its out‑links, multi‑hop chains are authored, and every non‑answer entity is rewritten into a descriptive reference so that no clue can be solved by simple string matching. Only questions that a reference model fails in closed‑book mode but solves once supporting evidence is provided are kept. These questions are turned into trajectories, which are filtered at both the trajectory and turn levels before supervised fine‑tuning (SFT).
The policy is then optimized by reinforcement learning (RL) against live search, with a reward judge and an observation summarizer running inside the training cluster. Over‑long rollouts are interrupted at the request level and resumed from their committed prefix in the next step. We alternate SFT and RL in a procedure called SFT‑RL climbing, feeding the hardest solved and most efficient rollouts of each RL round back into the next supervised pass.
Because inference‑time context management contributes more to these benchmarks than most reported system differences, we evaluate every benchmark both with and without it, keeping the tool set, context limit, and judge fixed. All results come from a single ReAct agent, with no sub‑agents and no test‑time verification. With management enabled, on BrowseComp, BrowseComp‑ZH, DeepSearchQA, and HLE the two models achieve $82.2/84.8/86.9/52.3$ and $88.6/85.1/92.9/56.4$, the strongest overall results among open‑source search agents in their respective parameter ranges. We plan to release the model weights together with the full recipe for data construction, training, and evaluation.
Review: The Iris series demonstrates a powerful combination of reverse‑task construction, rigorous trajectory filtering, and alternating SFT‑RL optimization, achieving state‑of‑the‑art performance on multilingual search benchmarks. The systematic study of context management also offers valuable guidance for future open‑source search research.