Iris-mini (35B-A3B) and Iris-pro (397B-A17B) use a specialized data pipeline that reverse-constructs multi-hop reasoning chains from web hyperlink structures. The training recipe utilizes RL optimized against live search to improve agentic navigation through descriptive reference rewriting.
HOW THIS AFFECTS YOU
●
builderYou can leverage these scale benchmarks to evaluate your own agentic reasoning capabilities.
●
researcherThe method of distilling multi-hop chains from entity graphs provides a new template for training search-specialized agents.