Iris: Climbing to the Search Frontier

Iris：攀登搜索前沿

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. 我们推出了 Iris-mini 和 Iris-pro，这是两个分别在 35B-A3B 和 397B-A17B 规模下训练的搜索智能体，并公开了其背后的数据流水线和训练方案。

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. 任务是根据网络语料库的超链接结构反向构建的：我们从种子页面及其外链中提取实体图，并在其上编写多跳链；将每个非答案实体改写为描述性引用，以确保无法通过字符串匹配解决线索；并且仅采纳那些参考模型在闭卷状态下无法回答、但在提供支持证据后能够解决的问题。

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. 这些问题随后被转化为轨迹，并在监督微调（SFT）之前在轨迹和轮次层面进行过滤。随后，策略通过针对实时搜索的强化学习（RL）进行优化，其中奖励判断器和观察总结器部署在训练集群内部，过长的执行过程会在请求层面被中断，并在下一步从已提交的前缀处恢复。

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. 我们交替进行这两个阶段，称之为“SFT-RL 攀登”过程，并将每一轮强化学习中解决难度最高且效率最高的轨迹反馈给下一次监督学习。

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. 由于在这些基准测试中，推理时的上下文管理比大多数系统间报告的差异更具价值，我们对每个基准测试都进行了有无上下文管理的评估，同时保持工具集、上下文限制和判断器不变。

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. 所有结果均来自单一的 ReAct 智能体，不包含子智能体，也不进行测试时验证。在开启上下文管理的情况下，在 BrowseComp、BrowseComp-ZH、DeepSearchQA 和 HLE 基准测试中，这两个模型分别达到了 82.2/84.8/86.9/52.3 和 88.6/85.1/92.9/56.4 的成绩，这是各自参数范围内开源搜索智能体中的最强表现。

We plan to release the model weights together with the complete recipe for data construction, training, and evaluation. 我们计划发布模型权重，以及完整的数据构建、训练和评估方案。