Iris: Climbing to the Search Frontier

Iris trains search agents on questions constructed to require finding connected evidence. Its pipeline builds question chains from linked webpages, replaces recognizable entity names with descriptions, and keeps questions a reference model cannot answer from memory but can solve when given supporting evidence. Successful search attempts then become training examples. The researchers alternate learning from those examples with reinforcement learning against live search, feeding difficult solved cases and efficient search sequences into the next training round. This creates a repeatable way to turn search experience into further instruction. The paper abstract describes the method and says model weights and the complete recipe are planned for release.