"Iris: Climbing to the Search Frontier" by Ziyuan Liu , Hengqi Liu , Zichuan Wang , Yang Qin , Jiachen Liang , Xu Chu , Shaowei Chen , Yuantao Gu , Mu Chuan
TLDR:
The text discusses the development of two advanced search agents, Iris-mini and Iris-pro, which were trained through a multi-stage pipeline involving supervised fine-tuning and reinforcement learning. These agents achieved top-notch performance on challenging web benchmarks by employing stringent trajectory filtering and context management during inference. The training process involved creating tasks from the hyperlink structure of a web corpus and refining questions to ensure they necessitated complex reasoning rather than simple string matching for resolution. The search agents were optimized using reinforcement learning against live search, with a strong emphasis on efficient trajectory filtering and management of context during inference, which proved crucial for high performance on various benchmarks. The results showed that these search agents outperformed others in the open-source domain within their respective parameter ranges. The intention is to release the model weights along with the full data construction, training, and evaluation recipe for interested users.
Read Paper / Blog
huggingface.co
Paper page - Iris: Climbing to the Search Frontier
Join the discussion on this paper page
1September 7, 2026 41
