Iris: Climbing to the Search Frontier (arxiv.org)

Iris-mini and Iris-pro, two state-of-the-art search agents, have been unveiled, trained at scales of 35B-A3B and 397B-A17B, respectively. This innovative approach to AI-driven search leverages a unique data pipeline where tasks are reverse-engineered from the underlying hyperlink structure of a web corpus. By generating multi-hop questions and applying techniques to prevent string-matching solutions, the models are tailored to excel in complex querying scenarios. The training process incorporates a method termed SFT-RL climbing, where reinforcement learning is optimized through interactions with live search scenarios, allowing models to evolve based on feedback from their performance. This announcement is significant for the AI/ML community as it sets new benchmarks in the performance of open-source search agents, showcasing the effectiveness of context management during inference—a critical factor in achieving superior results. Iris models reached impressive scores on multiple platforms, surpassing previous open-source results and indicating a substantial leap in the capability of AI search engines. The planned release of model weights and comprehensive training recipes will enhance the accessibility and replicability of this research, fostering further innovation in AI search technologies.