Iris: How SFT-RL Climbing Trains Open-Weight Agents to Search the Web Like a Researcher

Web search sounds simple until you try to automate it for hard questions. Ask a language model to find the founding date of a company, and it will usually succeed. Ask it to trace a multi-hop chain — "which researcher co-authored a 2019 paper with the person who later led the team that built X?" — and most agents fall apart within a few steps. They lose track of what they have already ruled out, exhaust their context window, or stop searching when the first plausible-sounding result appears.

Iris, a new open-weight search agent system from AllSpark Research, takes a systematic approach to this problem. The paper introduces two models — Iris-mini (35B total parameters, 3B active) and Iris-pro (397B total parameters, 17B active) — both post-trained from the Qwen3.5/3.6 MoE series. On benchmarks like BrowseComp and Humanity's Last Exam (HLE), they reach state-of-the-art results among open-source search agents. But the more interesting contribution is the training recipe and the inference-time insight that makes those numbers possible.

The Problem with Existing Search Agents

Most search agent research focuses on the policy — the model that decides what to search and how to interpret results. The execution harness (context management, episode restarting, trajectory filtering) is treated as an implementation detail. Iris argues this is backwards: for long-horizon search, the harness can matter as much as the policy itself.

The core challenge is context saturation. A multi-hop search task might require 20 or 30 tool calls before an answer emerges. Each call adds retrieved text to the context window. By the time the agent is halfway through a hard question, it may have consumed most of its 256K-token budget on intermediate evidence that is no longer relevant. Without a strategy for managing this, even a strong policy will fail — not because it reasoned poorly, but because it ran out of space to reason.

Building the Training Data from Web Graphs

Before training, the team needed hard, verifiable search tasks. Rather than hand-labeling questions, they built a fully automated pipeline that reverse-constructs tasks from the hyperlink structure of a web corpus.

The process works in three stages. First, the system creates a local subgraph from a seed page and its outbound links, then distills a compact entity graph. Second, it generates multi-hop questions where the answer requires traversing at least N hops through the entity graph. Third — and this is the key step — it applies an "anchor abstraction" operator that rewrites non-answer entities into descriptive references rather than proper names. This prevents the model from solving questions through simple string matching rather than genuine multi-step reasoning.

Tasks are only admitted to the training set if they pass dual-criteria verification: a reference model must fail them closed-book (confirming they are genuinely hard) but succeed when given the relevant evidence (confirming they are solvable). This filter keeps the training distribution challenging without making it impossible.

SFT-RL Climbing: An Iterative Training Loop

The training recipe, which the authors call SFT-RL climbing, alternates between supervised fine-tuning and reinforcement learning to continuously raise the difficulty ceiling.

The SFT phase trains on trajectories from a strong teacher model, filtered in two tiers. Coarse filtering removes trajectories that are incorrect, degenerate, or too shallow in search depth. Fine filtering uses an LLM judge to label individual turns, masking up to 10% of turns where the signal is noisy. This turn-level masking gives the model cleaner gradient signal than simply accepting or rejecting entire trajectories.

The RL phase optimizes the policy against live search environments using group-relative policy gradient. To handle long-horizon tasks, the system uses partial rollouts: when a trajectory exceeds the context limit, it is interrupted at the request level and the model resumes from its committed prefix rather than starting over. This allows the RL phase to explore trajectories that would otherwise be truncated.

The climbing part comes from feeding successful RL rollouts back into the SFT dataset. As the policy improves, the difficulty band shifts upward automatically — the model is always trained near the edge of its current capability, creating a self-reinforcing improvement cycle without manual curriculum design.

Inference-Time Context Management Is Not Optional

One of the paper's clearest findings is that inference-time context management (CM) is not a minor engineering concern — it is a first-class component of the search system.

The authors evaluate Iris under four settings: no context management, "retry" (restarting episodes that fail to produce a parseable answer while carrying forward a summary of ruled-out paths), "discard-all" (clearing tool history and restarting from the original question when context limits are reached), and a combination of both.

The performance differences are substantial. On BrowseComp, Iris-pro scores 88.6 with discard-all enabled, rising to 90.3 when combined with retry. Without any context management, the same model scores noticeably lower. For Iris-mini, the gap is even larger — smaller models with tighter effective context capacity benefit most from aggressive context clearing.

The practical implication is that evaluating search agents without specifying their context management strategy produces numbers that are not comparable across systems. A weaker policy with good context management can outperform a stronger policy running without it.

Benchmark Results

Evaluated on four benchmarks — BrowseComp, BrowseComp-ZH, DeepSearchQA, and the text-only subset of Humanity's Last Exam — Iris-mini and Iris-pro reach the following scores with the discard-all strategy:

Model	BrowseComp	BrowseComp-ZH	DeepSearchQA	HLE
Iris-mini (35B)	82.2	84.8	86.9	52.3
Iris-pro (397B)	88.6	85.1	92.9	56.4

These place both models at the top of open-source search agent rankings as of their September 2026 release. The models use a single ReAct-based architecture without sub-agents or test-time verification ensembles, keeping the inference setup straightforward to reproduce.

What This Means for Practitioners

Three takeaways stand out for anyone building or evaluating search-augmented LLM systems.

Context management deserves explicit design. Treating context handling as an afterthought likely leaves significant performance on the table. The discard-all strategy — clearing tool history when the context fills up — is simple to implement and consistently improves results.

SFT-RL climbing generalizes beyond search. The procedure — where RL discoveries feed back into SFT data and the difficulty band shifts upward automatically — is a general recipe for long-horizon agent training without manual curriculum design.

Open-weight models can compete on hard search tasks. Iris-pro at 397B parameters (17B active) reaches scores previously associated only with proprietary systems, and the post-training recipe is fully described in the paper.

The code, model weights, and datasets are available on the AllSpark Research GitHub repository.

Conclusion

Iris shows that the gap between open-weight and proprietary search agents is narrowing — not through architectural novelty, but through careful attention to training data quality, iterative policy improvement, and inference-time context handling. The SFT-RL climbing procedure is a reusable template for agents that need to sustain coherent reasoning across many steps and tool calls. The finding that context management is a first-class system component, not an implementation detail, is worth internalizing before your next search agent evaluation.