Paper Deep-Dive: Iris — Reverse-Built Search Tasks and SFT↔RL Climbing
A technical report that reverse-builds multi-hop questions from web link structure and alternates SFT with RL to train search agents — and that measures every benchmark both with and without context management, so policy strength can be told apart from harness tricks.
Key points
- Iris is the technical report for two search agents, Iris-mini (35B-A3B scale) and Iris-pro (397B-A17B scale).
- Training questions are reverse-constructed from the hyperlink structure of a web corpus, and every non-answer entity is rewritten into a descriptive reference.
- Only questions that a reference model fails closed-book, yet solves once the supporting evidence is supplied, are admitted.
- Trajectories are filtered before SFT, then the policy is optimized with RL against live search; alternating the two stages is called SFT-RL climbing.
- With context management enabled, BrowseComp scores are 82.2 (mini) and 88.6 (pro), from a single ReAct agent with no sub-agents and no test-time verification.
Iris: Climbing to the Search Frontier
Primary source — what this article is built on
undefined2026-09-03→undefined2026-09-10same month
Iris: Climbing to the Search FrontierZiyuan Liu, Hengqi Liu, Zichuan Wang et al. · 2026-09-03 · v1arXiv:2609.04304Paper page·PDFundefined
We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference so that no clue can be resolved by string matching, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at both the trajectory and the turn level before SFT. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster, and with over-long rollouts interrupted at the request level and resumed from their committed prefix at the next step. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark 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. With management enabled, on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE the two models reach $82.2/84.8/86.9/52.3$ and $88.6/85.1/92.9/56.4$, the strongest overall results among open-source search agents in their respective parameter ranges. We plan to release the model weights together with the complete recipe for data construction, training, and evaluation.
What the paper claims
The original title is "Iris: Climbing to the Search Frontier" (AllSpark Team, arXiv:2609.04304, published 2026-09-03). It is the technical report for two search agents, Iris-mini (35B-A3B scale) and Iris-pro (397B-A17B scale).
Here is the abstract in plain terms. Training questions are reverse-constructed from the hyperlink structure of a web corpus: every non-answer entity is rewritten into a descriptive reference so that no clue can be resolved by string matching. Only questions that a reference model fails closed-book, yet solves once the supporting evidence is supplied, are admitted. Those questions become trajectories, which are filtered at both the trajectory and the turn level before SFT; the policy is then optimized with RL against live search. Alternating the two stages is what the authors call SFT-RL climbing. And because inference-time context management is worth more on these benchmarks than most reported differences between systems, every benchmark is evaluated both with and without it. With management enabled, the numbers on BrowseComp / BrowseComp-ZH / DeepSearchQA / HLE are 82.2/84.8/86.9/52.3 (mini) and 88.6/85.1/92.9/56.4 (pro) — all from a single ReAct agent, with no sub-agents and no test-time verification.
The analogy: switch to writing the questions
The obvious way to get a stronger search agent is to reach for a stronger model. This paper starts somewhere else: first, manufacture a lot of genuinely hard questions. Naturally occurring web questions are mostly too easy, and a question whose answer sits on a single page will never teach an agent to walk across pages and reconcile clues (§2). So the authors draw the treasure map backwards. Fix the answer first, follow links out from its page to gather the neighbourhood, trace one path through the resulting web of relations, and turn the description of that path into the question.
The machinery: folding link structure into a relation graph
The corpus is modelled as a directed graph whose nodes are pages and whose edges are hyperlinks (§2.1). A seed page is drawn and its out-links are added to form a local subgraph — in effect, "one seed page plus the k pages it links to."
Raw pages carry too much noise, so the subgraph is distilled into an entity graph . The extractor keeps only entities and relations lying on a multi-hop path toward the seed theme, so what survives is a skeleton to write questions over. On that skeleton, the seed theme is fixed as the answer and a question is generated together with its reasoning path. Two constraints apply: the path must be at least hops long (so at least coupled relations are needed to solve it), and the answer must not appear in the question (§2.2).
Prying the clues loose from proper nouns
This is where this style of data synthesis earns its keep. If a proper noun survives in the question, the agent can skip the intended reasoning entirely and just throw that string into a search box. The paper kills the shortcut with anchor abstraction (§2.2).
In words: every entity other than the answer is replaced by a description that contains neither its name nor any alias — while still picking out exactly one referent. Delete "Isaac Newton," write "the seventeenth-century physicist who proposed the law of universal gravitation." Identical to a human reader; a completely different object to a search engine.
Why it is a different object becomes clear once you view retrieval as a question of proximity. Drag the query next to a rare proper noun (an isolated point with nothing around it), then drag it into a crowded region of near-duplicate descriptions, and watch how stable the top results are.
Comments
Sign in to comment