For the past six months, the search-agent race has centered on one question: how do you keep an agent from falling apart over many turns of live web search? Closed systems from OpenAI and Anthropic have pushed BrowseComp and HLE high, but the open-source side has never closed the gap — the base models are there, but no reproducible recipe has turned them into competent searchers. AllSpark Research's Iris, submitted to arXiv on 3 September 2026, attacks this from two sides at once: the training pipeline and inference-time context management.
Two sizes, one training recipe
Iris ships a pair of MoE models — Iris-mini at 35B total / 3B active parameters, Iris-pro at 397B total / 17B active — both numbers straight from the abstract. They share one training pipeline; only the base-model scale and post-training hyperparameters differ. A single data recipe scales cleanly across two deployment shapes.
The training procedure is what the authors call SFT-RL climbing: supervised fine-tuning to cold-start the policy, then reinforcement learning against live search APIs, alternating. During RL, the reward judge and the observation summarizer both run inside the training cluster to minimize reward-signal latency. For rollouts that exceed the length budget, the system interrupts at the request level and resumes from the committed prefix on the next step — a vent for the "latency trap" in long-context RL training.
The data-side key moves: reverse construction and reference rewriting
Iris's training tasks are not mined from existing QA corpora. They are reverse-engineered from the hyperlink structure of the web. Concretely:
- Take a seed page and fan out its outgoing links into an entity graph.
- Author multi-hop question chains over that graph.
- Rewrite every non-answer entity into a descriptive reference so no clue can be resolved by string match.
- Validate: the question must fail closed-book against a reference model, yet succeed once the supporting evidence is provided.
Step four is the quality gate — it filters out almost every "the answer is already in the question" sample. Surviving questions become trajectories, filtered at both trajectory and turn level before SFT. The three-stage pipeline — graph-based reverse mining, reference rewriting, closed-book failure validation — is an unusually engineering-grade answer to "the training question itself leaks the answer", a failure mode search agents are uniquely vulnerable to.
Benchmark numbers: four boards, two sizes, top of each band
The paper runs four benchmarks under one fixed setup — single ReAct agent, no sub-agents, no test-time verification, fixed tool set, fixed context window, fixed judge. With inference-time context management enabled:
- Iris-mini: BrowseComp 82.2 / BrowseComp-ZH 84.8 / DeepSearchQA 86.9 / HLE 52.3
- Iris-pro: BrowseComp 88.6 / BrowseComp-ZH 85.1 / DeepSearchQA 92.9 / HLE 56.4
The paper states explicitly that in their respective parameter bands, these are the strongest overall results among open-source search agents. On BrowseComp-ZH, pro (85.1) leads mini (84.8) by just 0.3 points, while the gap widens to 4+ points on the other three boards — a useful read on how scale interacts with depth of evidence retrieval vs. Chinese-language coverage.
The most consequential design choice in the benchmark section is the side-by-side reporting of "with / without context management" results, holding tool set, context window and judge constant. This separates "context management" from "the model itself" as independent variables — quantifying, in benchmark form, the claim that context management is worth more on these boards than most reported system-level differences. Engineering effort outside the training pipeline is here for the first time held up to the same scrutiny as the model itself.
The open release plan
Per the paper and the project page, AllSpark plans to ship together:
- Iris-mini and Iris-pro model weights (HF collection page: huggingface.co/collections/AllSpark-Research/iris);
- The full recipe for data construction, training, and evaluation;
- A companion GitHub repo at github.com/AllSpark-Research/Iris.
That means any team with 35B/400B-class compute can replicate the SFT-RL climbing pipeline on its own base model. This breaks from the past year's standard pattern of "open weights + open paper + locked training details" — how the training data was constructed is more worth publishing than how the training script was written.
Why this matters for search-agent training
Training-side: SFT-RL climbing is not new (the self-play lineage after RLHF is mature), but pairing it with reverse-engineered questions, reference rewriting and closed-book validation is what makes Iris's recipe defensible. The hidden cost of search-agent RL is reward hacking — any intermediate step can game "did the agent finally answer correctly" by leaking the answer. Iris turns "the question surface cannot contain the answer" into a hard constraint, halving the surface area available for hacking.
Inference-side: context management is not new vocabulary, but treating it as a peer-level experimental variable — rather than "something the base model already does" — is the methodological move worth copying. On high-difficulty boards like HLE, the relative lift from context management frequently exceeds the lift from swapping the base model — yet this used to be silently absorbed into the headline score.
Product-side: for any team shipping deep-research products, Iris's practical lesson is direct — stop chasing base-model upgrades, and instead build context management as a standalone module, harden tool invocation into a stable API, and tighten the reward judge into a low-latency service. These three efforts matter more to shipped quality than a single base-model bump.
The "so what": the open-source search-agent line has evolved from LangChain-era ReAct engineering, through WebGPT/WebShop-style "search + behavior" training signals, to Iris's "reverse-engineered questions + context management as a reproducible recipe" stage. The hard part of search-agent work is moving from "who can top the leaderboard" to "who can replicate the recipe". Iris's answer is the latter.
Sources: AllSpark Research; arXiv paper 2609.04304 (submitted 2026-09-03); #1 paper on Hugging Face Daily Papers (43 upvotes, topped 2026-09-07). Paper page: https://arxiv.org/abs/2609.04304 ; HF papers page: https://huggingface.co/papers/2609.04304 .