Search agents sit at an awkward intersection of language modelling and systems engineering. A model that can reason well in closed-book settings does not automatically transfer that capability to open-ended web interaction, where the agent must decide what to retrieve, how to interpret partially relevant results, and when it has gathered enough evidence to commit to an answer. The Iris paper (arXiv:2609.04304) addresses this directly, presenting an end-to-end recipe covering data construction, supervised training, reinforcement learning against live search, and a careful empirical treatment of context management. The result is two models, Iris-mini (35B) and Iris-pro (397B), that lead open-source search agent benchmarks across four challenging evaluations.
Key Contributions
The paper makes three substantive contributions that are worth separating out, because they each address a distinct failure mode in prior work.
- Task construction from web graph structure. Rather than relying on naturally occurring web questions, which are frequently too easy for training purposes, the authors reverse-construct multi-hop tasks from hyperlink graphs. Seed pages generate entity graphs; chains are drawn over those graphs; and then every non-answer entity is rewritten by an abstraction operator that strips away surface names and aliases while preserving unique identifiability. The result is questions that cannot be resolved by simple string matching against a search index. This is a meaningful step beyond prior entity-masking approaches, which often left enough lexical signal that a capable retriever could shortcut the intended reasoning chain.
- Iterative SFT and RL training. The training pipeline alternates between supervised fine-tuning on filtered teacher trajectories and reinforcement learning against live web search. Crucially, trajectories are filtered at both the trajectory level and the individual turn level before entering supervised training, and successful behaviours discovered during RL rollouts are consolidated back into the SFT stage before the next round of exploration begins. This iterative climbing structure is the source of the paper's title and is a sensible response to the instability that can arise when RL is applied directly to long-horizon interaction tasks without a strong behavioural prior.
- Transparent treatment of context management. This is arguably the paper's most useful methodological contribution to the field. Context management (CM) refers to inference-time strategies that handle the problem of long search trajectories exhausting the available context window before the agent has resolved all constraints. The authors evaluate their models with and without CM, report both sets of numbers, and analyse the gap. Many existing results in this space are reported only with CM enabled, making it genuinely difficult to attribute performance to the underlying policy versus the surrounding inference harness.
Methodology in Detail
The data pipeline deserves closer attention. Starting from a seed page, the system distills an entity graph from the page and its out-links, then authors multi-hop reasoning chains over that graph. The abstraction operator is defined formally: for every non-answer entity, the rewritten reference must not contain the entity's name or any known alias, yet must still uniquely identify it. This forces the agent to perform disambiguation through reasoning rather than retrieval by name. Questions that a reference model can answer in closed-book mode are discarded, as are questions the model cannot answer even when the supporting evidence is provided directly. This dual filter ensures the training set sits in a productive difficulty range.
For training, a strong teacher model generates trajectories over the filtered question set. These trajectories are filtered at the trajectory level (did the agent eventually reach a correct answer?) and at the turn level (were individual retrieval and reasoning steps well-formed?). The filtered trajectories seed the first SFT stage. RL then runs against live web search, with a judge and an observation summariser served within the training cluster to keep latency manageable. After each RL stage, successful new trajectories are harvested and folded back into the SFT corpus before the next round begins.
The CM analysis compares three strategies: no CM (the baseline that reflects the policy's intrinsic capability), discard-all (clearing the full interaction history and restarting when the context budget is nearly exhausted), and retry (summarising a failed attempt and appending it as prior knowledge for the next attempt). The authors also evaluate the combination of discard-all followed by retry.
Results
The quantitative results are strong. On BrowseComp, Iris-mini scores 82.2, outperforming the next-best model in the 30-35B range (XYZ-Aquila-mini at 78.8) by 3.4 points. Iris-pro reaches 88.6, a 3.8-point lead over XYZ-Aquila-pro. On Humanity's Last Exam, Iris-pro achieves 56.4, ahead of Kimi-K2.6 (1T parameters) at 54.0. These are meaningful margins, particularly given the parameter count comparisons involved.
The CM analysis is the more instructive set of results. Without any CM, Iris-mini scores 64.7 on BrowseComp; discard-all raises this to 82.2, a gain of 17.5 points. For Iris-pro, the same strategy yields a 16.0-point gain from a stronger 72.6 baseline. The authors make an important observation here: the smaller model benefits more from CM not because it has a larger context budget available, but because it consumes that budget faster, reaching the limit more often. CM is therefore most valuable for models that have already learned effective search behaviours but run out of room to execute them fully.
The benchmark-dependence of CM gains is also informative. BrowseComp, which requires sustained multi-hop retrieval and evidence integration, gains the most from context reset. HLE, where the primary challenge is expert-level knowledge and reasoning rather than search breadth, gains less. This makes sense: if the binding constraint is not context length but reasoning depth, resetting the context buys nothing.
Limitations and Broader Implications
The paper is candid about what it does not yet achieve. A gap remains relative to frontier systems such as Kimi-K3 and GPT-5.6 Sol, and closing that gap is left as future work. The BrowseComp-ZH results also show an interesting ceiling: three different configurations across two model scales all converge on exactly 85.1, suggesting the remaining errors on that benchmark may reflect data quality or question ambiguity rather than model capability. The authors flag this and discuss it in an appendix.
The inference cost of the retry strategy is also worth flagging. Combining discard-all with retry pushes Iris-pro above 90 on BrowseComp, but each retry requires a full additional search attempt. The authors are appropriately cautious about presenting this as a primary result, and their decision to report discard-all as the default in the main table reflects a reasonable position on the benchmark-versus-capability tradeoff.
Perhaps the most thought-provoking observation in the paper is the transfer result mentioned in the conclusion. Search-specialised training and search-derived teacher models transferred positively to General Tool Use and Cowork tasks that were not explicitly targeted. If this generalises, it suggests that search training may be better understood as instilling a general capability for acting under incomplete information, rather than as a narrow vertical specialisation. That framing has real implications for how agentic training pipelines should be structured, and it is a direction worth watching as the field matures.
The planned release of model weights and pipeline components is also significant. Much of the progress in search agents has been difficult to reproduce because the inference-time harness, not just the model, determines reported performance. Releasing both together would give the community a cleaner baseline from which to measure genuine progress.
Full paper: arXiv:2609.04304