Iris: Climbing to the Search Frontier
Iris reports open-source search-agent results built from a disclosed SFT-RL training loop and live-search evaluation.
The paper introduces Iris-mini and Iris-pro, trained at 35B-A3B and 397B-A17B scale. Its data pipeline builds multi-hop questions from web hyperlink graphs, filters them through evidence-based solvability checks, then trains search trajectories with supervised learning and reinforcement learning. With context management enabled, the models post top open-source results in their parameter ranges across BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE. The authors say they plan to release both model weights and the full construction, training, and evaluation recipe. HF Daily Papers' note
The paper introduces Iris-mini and Iris-pro, trained at 35B-A3B and 397B-A17B scale. Its data pipeline builds multi-hop questions from web hyperlink graphs, filters them through evidence-based solvability checks, then trains search trajectories with supervised learning and reinforcement learning. With context management enabled, the models post top open-source results in their parameter ranges across BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE. The authors say they plan to release both model weights and the full construction, training, and evaluation recipe. HF Daily Papers' note
score 6