Iris-pro 397B search agent hits 88.6% BrowseComp, 56.4% HLE

Iris-pro, an open-source search agent at the 397B-A17B scale, reaches 88.6 on BrowseComp and 56.4 on HLE, its authors report in a new arxiv paper. The smaller Iris-mini, at 35B-A3B, hits 82.2 and 52.3 on the same two benchmarks. On DeepSearchQA and BrowseComp-ZH the pair reach 86.9/92.9 and 84.8/85.1, which the paper calls "the strongest overall results among open-source search agents" in their respective parameter ranges.

One line in the abstract stands out. "Inference-time context management is worth more on these benchmarks than most reported differences between systems," the paper says. Every benchmark is therefore evaluated both with and without it, "holding the tool set, the context limit, and the judge fixed." All results come from "a single ReAct agent, with no sub-agents and no test-time verification."

The training recipe is unusual. Tasks are "reverse-constructed from the hyperlink structure of a web corpus": multi-hop chains over an entity graph distilled from a seed page and its out-links, with every non-answer entity rewritten "into a descriptive reference so that no clue can be resolved by string matching," admitting only questions "that a reference model fails closed-book yet solves once the supporting evidence is supplied."

The authors then alternate supervised fine-tuning with reinforcement learning against live search, in a procedure they call "SFT-RL climbing," returning "the hardest solved and most efficient rollouts of each RL round to the next supervised pass." Over-long rollouts are "interrupted at the request level and resumed from their committed prefix at the next step."

The abstract publishes no per-benchmark with-and-without numbers, does not compare Iris to closed frontier systems, and does not disclose training compute. The authors say they plan to release the model weights alongside "the complete recipe for data construction, training, and evaluation."