Source-linked AI summary
Iris: Climbing to the Search Frontier
Ziyuan Liu, Hengqi Liu, Zichuan Wang, Yang Qin, Jiachen Liang, Xu Chu, Shaowei Chen, Yuantao Gu, Mu Chuan
TL;DR
Search agents need to reason over retrieved evidence while managing long, dynamic trajectories. Iris addresses this with web-graph-derived multi-hop data, filtered SFT, live-search RL, and iterative SFT–RL climbing, achieving leading open-source results across four benchmarks while explicitly measuring context management.
Problem
Search agents must decide what to search, interpret evidence, continue exploration, and determine when available evidence is sufficient in dynamic environments.
Method
Iris reverse-constructs difficult multi-hop tasks from web graphs, filters trajectories and turns for SFT, then alternates live-search RL with supervised training through SFT–RL climbing.
Results
Iris-pro leads or matches the best result on all four benchmarks, including 88.6 on BrowseComp, 85.1 on BrowseComp-ZH, 92.9 F1 on DeepSearchQA, and 56.4 accuracy on HLE.
Takeaways & Limitations
The paper presents an end-to-end recipe for building search agents and reports strong open-source performance across four challenging benchmarks.
Takeaways & Limitations
Retry can improve scores but requires another full search attempt, so the reported primary configuration uses discard-all results rather than retry.
Abstract
from arXiv · showhide
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.
1 Introduction
Iris presents an end-to-end recipe for training strong search agents and evaluates them with and without context management. Its contributions combine difficult multi-hop task construction, filtered training trajectories, iterative SFT–RL training, and reproducible evaluation.
- Motivation: Context management is part of the effective inference system because long-horizon search can exhaust the available context before all constraints are resolved.The evaluation compares managed and unmanaged context while holding the tool interface, context budget, and judging procedure fixed.
- Contributions: The pipeline constructs challenging multi-hop search tasks from web structure and applies trajectory-level and turn-level filtering to obtain high-quality supervised data.The tasks remove easily searchable anchors and retain questions that are difficult yet objectively verifiable.
- Evaluation: Iris-mini and Iris-pro are evaluated across BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE under both context-management regimes.The setup is designed to make the effect of context management directly measurable.
- Contributions: Filtered SFT, live-web RL, and iterative SFT–RL climbing feed successful trajectories discovered during RL back into later supervised training rounds.This procedure combines the two training stages rather than treating them as a one-time sequence.
- Contributions: The authors plan to release model weights and key data-construction, training, and evaluation components to facilitate reproduction and further research.
2 Data Pipeline
The data pipeline reverse-constructs difficult, verifiable multi-hop questions from web-graph structure. It distills linked pages into an entity graph, abstracts searchable anchors, and keeps only questions that are hard closed-book yet solvable with supplied evidence.
- Motivation: The pipeline targets questions that require composing evidence across several sources and cannot be answered from parametric memory alone.It uses web-graph structure because naturally occurring questions rarely satisfy both requirements, while hand-written questions are expensive to scale.
- Web-graph construction: The corpus is modeled as a directed graph of pages and hyperlinks, from which a seed page is expanded along out-links into a local subgraph.The answer-anchored mode fixes a target answer entity before retrieving and expanding relevant pages.
- Entity-graph extraction: A subgraph is distilled into a compact entity graph whose salient entities and typed relations preserve multi-hop structure for precise question authoring.The extractor keeps entities and relations on paths toward the seed theme.
- Multi-hop question generation: Questions are generated with a reasoning path over the entity graph, and a hard path-length constraint requires dependence on at least N coupled relations.The seed theme is used as the target answer.
- Anchor abstraction: Anchor abstraction rewrites every non-answer entity into a descriptive reference, preserving the reasoning structure while preventing direct string matching.The resulting question requires disambiguation by reasoning rather than searching a surface string.
- Dual-criteria verification: A question–answer pair is retained only when a reference model fails it without tools but solves it when the graph evidence is supplied.Semantic matching determines answer equality, and the accepted set contains verified, unique answers.
3 Training Recipe
The training recipe combines filtered supervised trajectories, live-search reinforcement learning, and iterative SFT–RL climbing. It filters both complete rollouts and individual turns, manages long-horizon execution through prefix reuse, and feeds selected RL discoveries back into supervised training.
- Supervised fine-tuning: SFT begins from ReAct trajectories that interleave reasoning, SEARCH or SCRAPE calls, and observations before producing a final answer.Each observation is a document-level summary rather than a raw page, keeping trajectories within a bounded context budget.
- Supervised fine-tuning: Trajectory-level filtering admits only rollouts that terminate successfully, receive a correct judge verdict, avoid degeneracy, and contain sufficiently deep tool use.Detectors remove repetition loops, runaway calls, malformed thought blocks, and duplicate message sequences; shallow direct-lookups are also discarded.
- Supervised fine-tuning: Turn-level filtering removes locally poor reasoning or actions from otherwise acceptable trajectories before supervised training.This targets redundant searches, hallucinated tool names, and reasoning inconsistent with the action taken.
- Reinforcement learning: RL uses live search with group-relative policy gradients, while in-house engines provide reward judging and observation summarization inside the training cluster.This avoids external API dependence during training and keeps the rollout context aligned with the trained observation format.
- Reinforcement learning: Request-level partial rollouts interrupt over-long sessions and resume them from committed prefixes, reusing completed turns with truncated importance sampling.The synchronous schedule keeps rollout GPUs busy, with roughly 2× over-sampling used as headroom for unfinished sessions.
- Iterative climbing: Each climb alternates RL exploration with SFT on rare, successful, efficient rollouts selected from cases whose current-policy pass rate satisfies 0 < ¯R(q) ≤1/2.The resulting difficulty band shifts toward harder examples as the policy improves, while minimum depth and shortest-valid-trajectory selection discourage lucky or unnecessary search.
4 Evaluation
The evaluation compares Iris with and without context management across four benchmarks, using fixed tools, context limits, and judging procedures. Iris achieves strong results across parameter ranges, while context management improves performance most when searches exhaust the available context.
- Benchmarks: BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE evaluate entity identification, Chinese-source search, evidence completeness, and answer accuracy.DeepSearchQA uses F1, while the other three benchmarks use accuracy.
- Evaluation protocol: The reported results are single-rollout pass@1 evaluations using an LLM judge, with the same tool set, context limit, and maximum turn budget.Table 1 reports discard-all context management for the authors’ models.
- Main results: Iris-mini scores 82.2 on BrowseComp, 84.8 on BrowseComp-ZH, and 52.3 on HLE, leading its 30–35B parameter range on those benchmarks.Its DeepSearchQA F1 is 86.9, below XYZ-Aquila-mini’s 89.5.
- Main results: Iris-pro leads or matches the best result on all four benchmarks, scoring 88.6 on BrowseComp, 85.1 on BrowseComp-ZH, 92.9 on DeepSearchQA, and 56.4 on HLE.It ties XYZ-Aquila-pro at 85.1 on BrowseComp-ZH.
- Context management: Without context management, Iris-mini scores 64.7 on BrowseComp and 72.3 on BrowseComp-ZH, outperforming several smaller search agents; Iris-pro adds 7.9 and 4.5 points.The advantage remains after context management is introduced, indicating that the gains are not solely due to the inference-time strategy.
- Context management: Context management consistently improves the no-management baseline, with larger gains for Iris-mini than Iris-pro and up to 21.2 points on Iris-mini’s BrowseComp.The smaller model consumes more steps resolving the same constraints, so context management has more to recover.
- Context management: Combining discard-all with retry produces the strongest result in most settings but requires another full search attempt for each retry.The authors therefore treat retry as an exploration of an inference-time upper bound rather than their primary reporting configuration.
5 Conclusion: Beyond Search
The report presents Iris as an end-to-end search-agent recipe instantiated at two scales and identifies transfer beyond web search as a promising direction. Search data and search-specialized teachers also transferred positively to several general tool-use and cowork benchmarks.
- Conclusion: The recipe covers data construction, training, and evaluation and is instantiated in Iris-mini and Iris-pro.Training uses reverse-constructed hyperlink tasks, filtered teacher trajectories, supervised training, live-search reinforcement learning, and alternating stages.
- Beyond search: Both synthesized search data and search-specialized models transferred positively to General Tool Use and Cowork benchmarks not explicitly targeted.The reported domains include BFCL, τ-bench, OfficeQA, and APEX.
- Beyond search: The authors interpret these transfers as evidence that search may be an atomic capability whose induced behaviors are reusable under incomplete information.They propose studying search data and search-derived teachers throughout training rather than only in a separate specialization stage.
6 Contributions
The supplied contribution-section passages identify the paper’s authors and state that authors are listed in contribution order.
- Contributors: Authors are listed in order of contribution.The passage provides the author list beginning with Ziyuan Liu, Hengqi Liu, Zichuan Wang, and Yang Qin.
- Contributors: The author list includes Ziyuan Liu, Hengqi Liu, Zichuan Wang, Yang Qin, Jiachen Liang, and Xu Chu.These names appear in the supplied author listing.
- Contributors: The author list also includes Shaowei Chen, Yuantao Gu, Mu Chuan, and contribution markers attached to some names.The supplied listing marks Yang Qin and Mu Chuan with symbols.
A A Case of Ground-Truth Inconsistency
Question 85 exposes a mismatch between BrowseComp-ZH’s official ground truth and the source material: the benchmark credits “Lannister,” while the series supports “Bolton” as Sansa’s second marriage.
- Benchmark mismatch: The agent answered “Bolton,” but the official BrowseComp-ZH ground truth was “Lannister.”This disagreement caused the system to receive no credit for the question.
- Source-material resolution: The question’s clues identify Sansa Stark, whose first formal marriage was to Tyrion Lannister and second marriage to Ramsay Bolton.The passage distinguishes formal marriages from Sansa’s earlier betrothal to Joffrey Baratheon, which did not result in marriage.
- Implication: The case suggests an inconsistency between the benchmark annotation and the underlying source material.The authors connect this observation to the need for search benchmarks with higher-quality annotations and broader capability coverage.