Iris: Climbing to the Search Frontier
Source: HuggingFace Daily Papers, 2026-09-07 · arXiv 2609.04304 · raw
TL;DR. Two open search agents, Iris-mini at 35B-A3B and Iris-pro at 397B-A17B, released with the full data pipeline and training recipe. On BrowseComp, BrowseComp-ZH, DeepSearchQA and Humanity's Last Exam they reach 82.2 / 84.8 / 86.9 / 52.3 and 88.6 / 85.1 / 92.9 / 56.4, the strongest open-source results in their parameter classes, all from a single ReAct agent with no sub-agents and no test-time verification. The claim in the paper that matters most to this wiki is not the score. It is a methodological one, stated plainly: inference-time context management is worth more on these benchmarks than most reported differences between systems, so they evaluate every benchmark both with and without it, holding tool set, context limit and judge fixed.
The pieces worth stealing
Task construction from link structure, with string matching deliberately broken. Multi-hop chains are built over an entity graph distilled from a seed page and its out-links. Every non-answer entity is then rewritten into a descriptive reference, so no clue in the question can be resolved by matching a string against the corpus. Only questions that a reference model fails closed-book but solves once the supporting evidence is supplied are admitted, which filters out both trivia the model already knows and questions the evidence does not actually answer.
Two-level trajectory filtering before supervised fine-tuning. Trajectories are filtered at the trajectory level and again at the turn level, so a mostly-good rollout does not teach a bad turn.
RL against live search, with the judge and summarizer served inside the training cluster. The reward judge and the observation summarizer run in-cluster rather than as external API calls, and over-long rollouts are interrupted at the request level and resumed from their committed prefix at the next step rather than discarded. That is a prefix-reuse discipline applied to RL rollouts, and it is the same economics the KV cache page has been recording on the serving side.
SFT-RL climbing. The two stages alternate, and each RL round returns its hardest solved and most efficient rollouts to the next supervised pass. Difficulty and efficiency, not just correctness, are the selection criteria.
Relation to prior wiki state
The context-management disclosure is the important contribution and it is an indictment. ContextPipe (09-06) argued that context assembly is structurally the same problem as relational query execution and cut an agent's total token volume 31% and its call count 23% by planning the prompt instead of appending to it. Iris says the same lever, applied at inference time, is larger than the gaps between the systems on a public leaderboard. Put together: the search-agent leaderboard has been substantially measuring harness quality while reporting it as model quality, and Iris is the first entry to say so and publish both columns.
That is the agent harness engineering page's central thesis (the harness, not the model, is the primary cost and design surface) arriving as a leaderboard-integrity problem rather than as a cost argument. The page should now hold a stronger version: any agent benchmark that does not fix or disclose its context-management policy is not comparing models.
The no-sub-agents result is a quiet counterpoint to the day's other agent signal. The DAIR.AI weekly digest surfaced Harness-of-Harness, which wraps an existing coding harness in repeated plan-code-test increments and reports a 52.25% average relative gain over standalone harnesses after three iterations. Iris reaches the open-source frontier on search with one flat ReAct loop. Both cannot be the general rule, and the honest reading is that the orchestration premium depends on whether the task decomposes: multi-day software projects do, single-answer multi-hop retrieval does not.
Gaps
The headline numbers are quoted with context management enabled, and while the paper commits to reporting both, the abstract does not give the without-management column, so the size of the effect it identifies as decisive is not visible from the summary. No inference cost per query is reported for either model, which for a system whose main methodological claim is about inference-time technique is the omission that matters. And weights and recipe are promised rather than released.