SearchArt: Training Long-Horizon Search Agent with Scalable Synthetic and Verified Task
Lang Mei, Xiaohan Yu, Chong Chen, Liyan Liu, Xiangnan Chen, Jinchao Ma, Chao Feng, Li Huang, Siyu Mo, Sichen Kang, Yunkun Xu, Zhihan Yang, Zhujun Xue, Jingren Zhang, Qing He, Yingdi Huang, Hao Jiang, Ziao Ma, Zewei Pan, Minhao Sun, Zhuo Tao, Jinzhao Xiao, Gangtao Xin, Huanyao Zhang, Wenjian Zhang, Jiangshan Zhang, Guojie Zhu, Jiaxin Mao, Wentao Zhang
cs.IR, cs.LG
2026-07-25
Huawei's SearchArt trains a 27B search agent with a synthesize-and-verify task pipeline plus SFT/RL; 74.39 on BrowseComp-ZH nears Gemini 3.1 Pro, +13% average over base.
Turning large language models into search agents that autonomously finish long-horizon, multi-retrieval, multi-step tasks (deep search, deep research) is a crowded direction. But training these agents hits two walls: there is no scalable source of long-horizon task data, because human annotation cannot keep up; and the correctness of intermediate reasoning and tool calls is hard to evaluate and correct.
Huawei Cloud's SearchArt offers a synthesize-verify-then-post-train pipeline, aiming to approach closed-source frontier systems on deep-search and deep-research benchmarks with only 27B parameters.
Task synthesis comes in several flavors: information-seeking QA and matching search trajectories synthesized from web documents and automatically built evidence graphs; high-quality surveys and research reports reverse-engineered into a question (the long document is the ready-made answer); plus conversational and broad-retrieval questions that resemble real users. Difficulty is ranked by the topological complexity and evidence dispersion of the subgraph, keeping only the hardest instances.
The verification pipeline is the gatekeeper: each synthesized QA and trajectory must pass three checks, QA consistency, trajectory quality, and relevance of retrieved evidence. Only survivors enter the training set.
Post-training has two stages: SFT teaches correct tool use and workflow; RL refines the policy. Rewards are hybrid: an outcome reward judges the final answer (rules first, a judge model as fallback), process rewards police format and per-turn reasoning budget, and a self-evolving judge auditor hunts reward hacking.
The base is Qwen3.5-27B. After post-training the average relative gain across five benchmarks is 13.16%: BrowseComp 61.0 to 70.06, BrowseComp-ZH 62.1 to 74.39, BrowseComp-Plus 58.43 to 63.49, WideSearch 61.1 to 64.0, DeepResearch-Bench 44.7 to 52.55.
The standout is Chinese BrowseComp-ZH at 74.39, within 0.4 of Gemini 3.1 Pro (74.8) and 0.9 of GPT-5.4 (75.3). Among 30B-class open search agents (MiroThinker, OpenSeeker, RedSearcher and the rest) it is clearly the strongest.
| Benchmark | Base Qwen3.5-27B | SearchArt-27B |
| BrowseComp | 61.0 | 70.06 |
| BrowseComp-ZH | 62.1 | 74.39 |
| BrowseComp-Plus | 58.43 | 63.49 |
| DeepResearch-Bench | 44.7 | 52.55 |
On the English flagship BrowseComp, SearchArt's 70.06 still trails the closed-source leaders by a wide margin (Gemini 3.1 Pro 85.9, GPT-5.5 84.4); do not be misled by "matches closed-source" framing. What actually holds is two things: it is far ahead among same-size small models, and on Chinese and deep-research benchmarks a 27B open model genuinely reaches the lower edge of the closed-source frontier. For teams building their own search agents, the synthesize-and-verify pipeline is more worth copying than any headline score.
The large gap on English BrowseComp shows that pure-English, broad-fact hard retrieval is not closed by a small model plus post-training against closed-source compute. Many numbers in the paper are marked with an asterisk (self-reproduced) and are not strictly comparable to official figures. Task synthesis leans on strong teacher models (Qwen3.5-397B and the like) to generate trajectories, and the verification pipeline itself depends on judge models, so the "evaluation is the same problem as retrieval" loop is not truly broken, only eased with engineering.