The announcement covers two search agents, Iris-mini and Iris-pro, trained at large scales with a specific data pipeline and training approach. Tasks are reverse-constructed from web hyperlink structures, with questions designed to challenge models' reasoning abilities.
The training process involves multi-hop chain authoring, rewriting non-answer entities, and filtering trajectories before supervised fine-tuning. Reinforcement learning optimizes the policy against live search, with a focus on returning the hardest and most efficient rollouts for subsequent supervised passes.
Evaluation includes both with and without inference-time context management, emphasizing the importance of context handling in search benchmarks. The models achieve strong results across multiple datasets, outperforming other open-source search agents in their parameter ranges.
The release will include model weights and the complete data construction, training, and evaluation recipes, supporting reproducibility and further research.
Source: https://arxiv.org/abs/2609.04304