Iris: Climbing to the Search Frontier

Iris. Climbing to the Search Frontier. By Ziyuan Liu and eight others. Published on the third of September. Its code repository is worked out: about fifty-five source files, with a readme.

Search agents are trained on questions, and the questions are the bottleneck. A question taken off the web may be answerable from memory, or answerable by matching a string, and neither of those teaches an agent to search. Iris therefore starts from how the tasks are built rather than from the model. Two search agents are presented, Iris-mini and Iris-pro, trained at the thirty-five billion and three hundred and ninety-seven billion parameter scales, and released alongside them are the data pipeline and the training recipe that produced them.

Task construction is the substance of the contribution. Tasks are reverse-constructed from the hyperlink structure of a web corpus. An entity graph is distilled from a seed page and its out-links, and multi-hop chains are authored over that graph. Every entity that is not the answer is then rewritten into a descriptive reference, so no clue in the question can be resolved by string matching. Then comes an admission test: a question is kept only if a reference model fails it closed-book and solves it once the supporting evidence has been supplied. Those two conditions together guarantee the question needs searching and can be answered by searching.

Training runs in two stages that feed each other. Questions become trajectories, and those trajectories are filtered before SFT at both the trajectory level and the turn level. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster rather than called out to. Over-long rollouts are interrupted at the request level and resumed from their committed prefix at the next step, instead of being thrown away. The two stages then alternate, in a procedure the authors call SFT-RL climbing: the hardest solved and most efficient rollouts of each RL round are returned to the next supervised pass.

The evaluation makes an unusual choice and says why. Inference-time context management is more valuable on these benchmarks than most reported differences between systems, so every benchmark is evaluated both with it 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, across BrowseComp, BrowseComp-ZH, DeepSearchQA and HLE, the smaller model scores eighty-two point two on the first and fifty-two point three on the last, while the larger reaches eighty-eight point six and fifty-six point four on those same two. The two benchmarks in between land in a similar band for both. These are the strongest overall results among open-source search agents in their respective parameter ranges.

Two claims are being made at once here, and they come apart cleanly. One is that a data pipeline built to exclude memorised and string-matchable questions produces agents that search better. The other is that comparisons in this area have been confounded by inference-time context management, which is why both settings are reported for every benchmark rather than the flattering one. The authors plan to release the model weights together with the complete recipe for data construction, training and evaluation.

Iris. Climbing to the Search Frontier. By Ziyuan Liu and eight others. Published on the third of September. This is the long version: the whole paper rather than the abstract.

There is a measurement problem sitting underneath the search-agent literature, and this report takes it seriously enough to reorganise its own evaluation around it. Substantially different reported performance can arise from differences in the inference-time harness rather than from differences in the underlying policy. The particular culprit is context management. Long-horizon search can exhaust the available context before the agent has resolved every constraint, so the effective search budget is much smaller than the nominal context window, and systems paper over that with trajectory summarisation, state compression, selective history removal, or full resets. Report only the managed number and you cannot tell how much came from the agent and how much from the wrapper around it.

Before any of that, the training data. A useful question has to be unanswerable from parametric memory and has to require composing evidence from several sources, and naturally occurring web questions rarely satisfy both. So questions are reverse-constructed from the hyperlink structure of a web corpus: a seed page is drawn in an answer-anchored mode that fixes the target entity first, the seed is expanded along its out-links into a local subgraph, and multi-hop chains are authored over the entity graph that results. Every entity that is not the answer is then rewritten into a descriptive reference, so no clue in the question can be resolved by string matching.

Training alternates two stages. Trajectories from a strong teacher are filtered twice before supervised fine-tuning, first at the trajectory level for correctness, degeneracy and search depth, then at the turn level by a judge whose rubric is induced from the data rather than written by hand. The policy is then optimised by reinforcement learning against live search, with the reward judge and the observation summariser served inside the training cluster rather than called out to, and with over-long rollouts interrupted at the request level and resumed from their committed prefix instead of discarded. The two stages then alternate, in what the authors call climbing: the hardest solved and most efficient rollouts of each reinforcement round are returned to the next supervised pass.

Now the measurement. Every benchmark is evaluated twice, with context management and without it, holding the tool set, the context limit and the judge fixed, and all results come from a single agent with no sub-agents and no test-time verification. The managed numbers are strong: the two models reach eighty-two point two and eighty-eight point six on BrowseComp. The unmanaged numbers are the interesting ones. The smaller model scores sixty-four point seven on that same benchmark with no context management at all. That gap, roughly seventeen points on one model from the harness alone, is larger than most of the differences between systems that the field reports as progress.

The comparison is bounded in ways the report states. It covers open-source search agents within matched parameter ranges, not the frontier generally. Some inference-time mechanisms are deliberately excluded from the definition of context management, including verification and re-answering strategies other systems use, on the grounds that those are additional test-time machinery rather than management. The authors describe this as an initial exploration focused on widely recognised search benchmarks, leaving a broader space of agentic scenarios untouched. And the unmanaged comparison is only possible against the handful of prior systems that reported unmanaged numbers at all.

The observation the authors close on points away from search entirely. Both the synthesised search data and the search-specialised models, used as teachers for distillation, transferred positively to domains nobody targeted: general tool use and office-style collaborative work. Their reading is that search may be better understood as an atomic capability than as a vertical specialisation, because the behaviours it induces are not really about browsing. They are about acting under incomplete information, which is most of what an agent ever does.