Iris：攀登搜尋前沿

摘要

我們提出 Iris-mini 與 Iris-pro 兩個搜尋智能體，分別在 35B-A3B 與 397B-A17B 規模上訓練，並介紹它們背後的資料管線與訓練配方。任務是從網路語料庫的超連結結構反向建構而來的：我們在從種子頁面及其外連提煉出的實體圖譜上構造多跳鏈，將每個非答案實體改寫為描述性指稱，使任何線索都無法經由字串比對解出；我們僅納入參考模型在閉卷條件下無法回答、但一旦提供佐證證據即可解出的問題。這些問題接著被轉換為軌跡，並在 SFT 之前同時於軌跡層級與輪次層級進行過濾。策略隨後以 RL 對即時搜尋進行最佳化；獎勵評判器與觀察摘要器以服務形式部署於訓練叢集內，過長的 rollout 會在請求層級被中斷，並在下一步從已提交的前綴繼續。我們以一種稱為 SFT-RL climbing（SFT-RL 攀升）的程序交替兩個階段，將每一輪 RL 中難度最高且成功解出的 rollout，以及效率最高的 rollout，回傳至下一個監督式階段。由於在這些基準上，推論期間的上下文管理比多數系統之間已發表的差異更具影響力，我們固定工具集、上下文長度限制與評判器，並對每個基準分別評估啟用與未啟用該機制的表現。所有結果皆來自單一 ReAct 智能體，沒有子智能體，也沒有測試時驗證。在啟用上下文管理的情況下，兩個模型在 BrowseComp、BrowseComp-ZH、DeepSearchQA 與 HLE 上分別達到 82.2/84.8/86.9/52.3 與 88.6/85.1/92.9/56.4，是各自參數範圍內開源搜尋智能體中整體表現最強的結果。我們計畫釋出模型權重，以及資料建構、訓練與評估的完整配方。

English

We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference so that no clue can be resolved by string matching, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at both the trajectory and the turn level before SFT. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster, and with over-long rollouts interrupted at the request level and resumed from their committed prefix at the next step. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark both with and without it, holding the tool set, the context limit, and the 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 reach 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 complete recipe for data construction, training, and evaluation.