Hugging Face has introduced Iris-mini and Iris-pro, two open-source search agents designed to tackle complex web-based tasks. The models, built at scales of 35B-A3B and 397B-A17B respectively, are trained using a multi-stage pipeline that combines supervised fine-tuning (SFT) with reinforcement learning (RL) against live search.
The training data is constructed by reverse-engineering questions from the hyperlink structure of a web corpus. The process involves creating multi-hop chains over an entity graph, rewriting non-answer entities into descriptive references to prevent trivial string matching, and filtering questions that a reference model cannot answer without supporting evidence. These questions are then transformed into trajectories, which undergo rigorous filtering at both trajectory and turn levels before SFT.
After SFT, the policy is optimized via RL with live search, where the reward judge and observation summarizer run inside the training cluster. Over-long rollouts are interrupted and resumed from their committed prefix. The two stages alternate in a procedure called SFT-RL climbing, where the hardest solved and most efficient rollouts from each RL round are fed back into the next supervised pass.
The researchers emphasize the importance of inference-time context management, which they evaluate on every benchmark 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 or test-time verification. With management enabled, the models achieve strong scores on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE, marking the best overall results among open-source search agents in their parameter ranges.
The team plans to release the model weights along with the complete recipe for data construction, training, and evaluation. For questions or collaboration, they can be contacted via email or WeChat.