Found first: a primary source the press has not covered yet.
A paper submitted to arXiv on September 3, 2026 introduces two search agents, Iris-mini (35B-A3B) and Iris-pro (397B-A17B), trained with a method the authors call "SFT-RL climbing." The paper reports 88.6% on BrowseComp and 56.4% on HLE for Iris-pro, with context management enabled. Iris-mini reaches 82.2% on BrowseComp and 52.3% on HLE under the same conditions.
What the source says
The work is led by Ziyuan Liu with eight co-authors. The training recipe alternates supervised fine-tuning and reinforcement learning stages against a live search environment. Rather than curated question-answer datasets, tasks are reverse-constructed from web corpus hyperlink structures, building multi-hop entity chains over a knowledge graph in a way the authors say prevents string-matching shortcuts. Iris-pro scores 92.9% on DeepSearchQA and 85.1% on BrowseComp-ZH; Iris-mini scores 86.9% and 84.8% on the same benchmarks. Each model runs as a single ReAct loop, with no sub-agents and no test-time verification pass. All scores are reported both with and without context management, making that component's contribution separable from the model itself.
Why it matters
The authors' central empirical claim is that inference-time context management accounts for more variance in these benchmark scores than most of the model-level differences reported across systems in the field. That finding, if it holds, reframes where the real gains in search agent performance are coming from. The task construction approach is also worth attention on its own terms: by mining hyperlink graphs rather than curating question sets, the method reduces the contamination risk that shadows most QA benchmarks. The paper leaves open how the scores compare to proprietary systems under equivalent conditions, but the benchmark numbers and the training recipe are fully described.