Iris trains 35B and 397B search agents by reverse-constructing questions from hyperlink structure, and reports benchmarks both with and without context management
arXiv 2609.04304 (2026-09-03, 47 HuggingFace upvotes) releases Iris-mini (35B-A3B) and Iris-pro (397B-A17B) along with the data pipeline. Tasks are built backwards from a web corpus's link graph: multi-hop chains over an entity graph distilled from a seed page and its out-links, with every non-answer entity rewritten as a descriptive reference so no clue survives string matching, admitting only questions a reference model fails closed-book but solves once given evidence. Training alternates SFT and RL against live search in a loop they call SFT-RL climbing, with the reward judge and observation summarizer served inside the training cluster. The methodological note is the one worth stealing: they state that inference-time context management is worth more on these benchmarks than most reported differences between systems, so they report every benchmark both with and without it, holding tools, context limit and judge fixed.
↳ Follow the thread