Iris: Climbing to the Search Frontier

AI summary

Overview

Research area: Artificial intelligence — training and evaluation of LLM-based web search agents (agentic retrieval, long-horizon tool use, reinforcement learning).

Technical level: Advanced. The paper assumes familiarity with supervised fine-tuning, group-relative policy gradient RL, ReAct-style agents, mixture-of-experts models, and inference-time context management.

Scope: A single paper from the AllSpark Team presenting Iris-mini and Iris-pro, two search agents at the 35B-A3B and 397B-A17B scales, together with the full data-construction, training, and evaluation recipe behind them.

What This Paper Is About

Existing web questions are usually too easy for training search agents, and hand-writing hard multi-hop questions does not scale. This paper builds hard questions automatically by walking the hyperlink structure of a web corpus, obscuring every entity so the answer cannot be found by string matching, and keeping only questions that a reference model fails from memory but solves when given the supporting evidence. It then trains search agents on these questions using an alternating SFT–RL loop, and reports results on four search benchmarks both with and without inference-time context management.

Key Contributions

1. An end-to-end data pipeline that reverse-constructs difficult, verifiable multi-hop search questions from the hyperlink structure of a web corpus, removing directly searchable anchors through entity rewriting and admitting only questions that a reference model fails closed-book yet solves when the evidence is supplied.
2. A multi-stage filtering scheme applied at two levels: trajectory-level filtering for correctness, degeneracy (via a sliding-window compression ratio and auxiliary detectors) and search depth, plus turn-level filtering by a judge whose rubric is induced from the data rather than written by hand (masking at most 10% of assistant turns).
3. An iterative SFT–RL climbing procedure in which RL runs against live search with an in-cluster reward judge and observation summarizer, and the hardest solved and most efficient rollouts of each round are returned to the next supervised pass.
4. A paired evaluation protocol reporting every benchmark both with and without context management, holding the tool set, context limit, and judge fixed, so that gains from the policy can be separated from gains from the inference-time harness.

Main Findings

• Headline results with context management enabled: Iris-mini reaches 82.2 / 84.8 / 86.9 / 52.3 on BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam; Iris-pro reaches 88.6 / 85.1 / 92.9 / 56.4. The authors describe these as the strongest overall results among open-source search agents in their respective parameter ranges.
• Iris-mini beats the strongest same-range model on three benchmarks: on BrowseComp it exceeds XYZ-Aquila-mini (78.8) by 3.4 points (82.2 vs. 78.8), and it also leads on BrowseComp-ZH (84.8) and HLE (52.3). On DeepSearchQA its F1 of 86.9 remains below XYZ-Aquila-mini's 89.5.
• Iris-pro leads or ties on all four benchmarks: highest BrowseComp score (88.6), outperforming XYZ-Aquila-pro (84.8) by 3.8 points; highest DeepSearchQA F1 (92.9); highest HLE accuracy (56.4), ahead of XYZ-Aquila-pro (53.3) by 3.1 points; and it ties XYZ-Aquila-pro at 85.1 on BrowseComp-ZH.
• A gap to frontier systems remains: heavier or larger systems report higher numbers in places, such as Kimi-K3 (2.8T) at 91.2 on BrowseComp and 95.0 on DeepSearchQA, Apodex-1.0-H at 90.3 on BrowseComp and 94.4 on DeepSearchQA, and Claude Fable 5 at 64.5 on the full HLE set. The authors state closing this gap is future work.
• Context management matters more than most reported differences between systems. For Iris-mini, discard-all is worth up to 21.2 points on BrowseComp, against at most 12.8 on BrowseComp-ZH, 9.2 on HLE, and 8.9 on DeepSearchQA. Gains are larger for Iris-mini than for Iris-pro.
• The gains are benchmark-dependent, not budget-dependent. HLE has the lowest no-context baseline of the four (43.2 for Iris-mini) yet gains less than BrowseComp; the ordering follows how often a session actually runs out of context, not how much headroom a benchmark leaves.
• Scores without any context management are also reported: Iris-mini reaches 64.7 on BrowseComp and 72.3 on BrowseComp-ZH without CM, ahead of FORT-Searcher (55.9 and 62.1), OpenSeeker-v2 (46.0 and 58.1), and REDSearcher (42.1 and 49.8). Iris-pro raises these two scores by a further 7.9 and 4.5 points respectively.
• The retry strategy gives the highest numbers: discard-all combined with retry pushes Iris-pro above 90 on BrowseComp (90.3), and the authors treat it as an exploration of the achievable upper bound rather than the primary reported configuration, because each retry costs another full search attempt.
• Three configurations converge on BrowseComp-ZH: Iris-mini under discard-all with retry, and Iris-pro under both discard-all and discard-all with retry, all score 85.1 — on a benchmark of 289 questions, 246 correct answers in all three cases, which the authors read as evidence that remaining performance is constrained by factors beyond model capacity.
• A ground-truth inconsistency is documented: in Question 85 of BrowseComp-ZH, the official ground truth is "Lannister" while the agent returns "Bolton"; the paper traces the events of Game of Thrones (Sansa Stark's first formal marriage to Tyrion Lannister in Season 3, her later marriage to Ramsay Bolton in Season 5) to argue the agent's answer is directly supported by the source material.
• Positive transfer beyond search: the synthesized search data and the search-specialized models used as teachers for on-policy distillation transferred to untargeted domains, including General Tool Use (BFCL and τ-bench) and Cowork (OfficeQA and APEX).

Methodology in Plain English

Building the questions. The corpus is treated as a directed graph where pages are nodes and hyperlinks are edges. Starting from a seed page (chosen in an answer-anchored mode that fixes a target answer entity), the pipeline expands along out-links into a local subgraph, recovering the true out-link set by merging a structured semantic mirror of the corpus (RDF triples) with rendered page markup. That subgraph is distilled into a compact entity graph of salient entities and typed relations lying on a multi-hop path toward the seed theme. A question is then authored over this graph together with its reasoning path, constrained to require at least N coupled relations and to exclude the answer. Every non-answer entity is rewritten into a descriptive reference that does not contain its name or alias but still uniquely identifies it, so no clue can be resolved by string matching.

Verification. A pair is kept only if a reference model fails the question closed-book (no tools) and gets it right when the entity graph is supplied as context. Answer equality is decided by a semantic-matching judge.

Making training data. A strong teacher solves each accepted question in the ReAct paradigm — interleaving reasoning, tool calls, and observations over live search tools (search and scrape) — with each observation being a document-level summary produced on the fly rather than a raw page. Trajectories must pass a trajectory-level gate: successful termination, correct answer under an LLM judge, no degeneracy, and at least K tool-call turns. Degeneracy is caught mainly by a sliding-window compression ratio (repetitive text compresses more than fluent text), supplemented by detectors for repeating lines, long single-character runs, and consecutive identical tool calls. An exact-duplicate pass then drops byte-identical trajectories.

Turn-level filtering. A judge labels individual assistant turns as keep or mask, seeing the question, reference answer, and a fixed local window of surrounding turns. The rubric is not hand-written: trajectories are sampled, critiqued in free form, and the recurring failure modes are consolidated into the rubric. Masked turns stay in the context but are excluded from the loss, and at most 10% of assistant turns in any trajectory are masked.

Training. SFT maximizes the teacher's output likelihood with the per-turn mask applied, using a replay operator to reconstruct the exact visible history the agent conditioned on. RL then optimizes against live search with a group-relative policy gradient. Long rollouts are handled by request-level partial rollout: once a step has committed enough completed trajectories, in-flight over-sampled sessions are aborted and resumed at the next step from their committed prefix, with truncated importance sampling correcting the mismatch between prefixes generated under different policy weights — at the cost of roughly 2× over-sampling as headroom. Rather than call an external API, several FP8 engines of an in-house Qwen3.5-397B-A17B model run inside the cluster and serve both as the generative reward model (a binary verdict against the reference answer, with no additive format term) and as the observation summarizer. There is no message-history pruning or sliding window in the rollout, so the context the policy conditions on is exactly what is trained on.

Climbing. SFT and RL alternate. For each query, at most one rollout from the existing RL samples is retained; queries with a pass rate above 0 and at most 1/2 are selected, and among successful rollouts with at least K_rft tool-call turns the shortest valid trajectory is chosen. Because the difficulty band is defined by the current policy's pass rate, it shifts toward harder examples as the policy improves.

Setup details. Iris-mini and Iris-pro are initialized from Qwen3.6-35B-A3B and Qwen3.5-397B-A17B, both mixture-of-experts with a 256K-token context window. SFT runs two epochs with a global batch size of 64 and a maximum sequence length of 262,144 tokens. RL uses the open-source Relax framework. Access to the Hugging Face dataset and Space pages hosting benchmark questions and answers is blocked at three points: removed from search results, refused on scrape, and caught by a post-hoc guard even when the model supplies the URL from memory. All results come from a single ReAct agent, with no sub-agents and no test-time verification. Evaluation is a single rollout (pass@1) per question, judged with each benchmark's official evaluation prompt; DeepSearchQA is scored with F1, while BrowseComp, BrowseComp-ZH, and HLE use accuracy. HLE is evaluated on the text-only subset unless marked otherwise.

Context management strategies compared. The main results use the discard-all strategy of DeepSeek-V3.2, which clears the entire interaction history and restarts the question when the running context reaches a threshold before a final answer is obtained. The retry strategy from MiroThinker is also evaluated; it summarizes a failed attempt into a compact description of unsuccessful search experience and appends it to the task for the next attempt. The authors deliberately do not use the reverify and reanswer procedures adopted by some systems, as those introduce test-time mechanisms beyond context management.

Why This Matters

The paper argues that reported differences between search agents can come from the inference-time harness rather than the underlying policy, and shows a way to report both regimes so the two can be told apart. It also supplies a reproducible recipe — data construction, training, and evaluation — and states an intent to release model weights.

Real-world applications:

• Long-horizon research assistants that must combine evidence scattered across many sources rather than retrieve a single answer span.
• Enterprise and domain search where the query is deliberately under-specified and the agent must disambiguate by reasoning rather than by keyword matching.
• General tool-using agents, since the authors report positive transfer of search data and search-derived teachers to General Tool Use and Cowork settings.
• Benchmark design and auditing, given the documented ground-truth inconsistency in BrowseComp-ZH Question 85.

Industry relevance: The paper's resource economics are notable for practitioners — serving the reward judge and observation summarizer in-cluster removes external API dependence from the training loop, and request-level partial rollout keeps rollout GPUs busy under a synchronous schedule. The finding that context management can be worth up to 21.2 points on BrowseComp, and that the benefit depends on how fast a model consumes context rather than on model capacity alone, directly affects how serving stacks for search agents should be built and how their results should be compared.

Future Directions

• Closing the gap to frontier systems. The authors state that a gap relative to the most capable frontier systems remains and that closing it is an important direction.
• New benchmarks with better annotations. The BrowseComp-ZH case study motivates developing new search benchmarks with higher-quality annotations and broader, more reliable capability coverage.
• Evaluation beyond benchmark scores. The paper argues that benchmark performance captures only part of an agent's capability and calls for evaluations in a broader range of realistic search scenarios.
• Search as an atomic capability rather than a vertical specialization. If the observed transfer to General Tool Use and Cowork generalizes, search data and search-derived teachers could be useful throughout training rather than confined to a separate specialization stage, with value measured by broader agentic competence — a direction the authors intend to pursue.

Target Audience

Researchers and engineers building search agents or agentic RL training pipelines; teams designing synthetic multi-hop question data; and practitioners who need to interpret published search-agent benchmark numbers, since the with-and-without context management protocol is the paper's most transferable methodological point. Readers without a background in reinforcement learning or agentic evaluation will find the training sections demanding, and the paper assumes prior familiarity with ReAct-style agents and mixture-of-experts models.

Authors’ abstract

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.