Xiaohongshu's AllSpark team has released the Iris Search Agent. The team has published the model weights and evaluation code, and plans to further open-source its data construction, training, and evaluation pipelines.

A Search Agent has to search, read, reason, and make decisions in a live web environment. It must judge what to search for and how to interpret retrieved results, and also decide when to keep searching and when to stop. The AllSpark team evaluated Iris-mini and Iris-pro on four benchmarks: BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam (HLE).

Under the evaluation setup published in the paper, Iris-mini posted the top score on three of the benchmarks among systems in the 30B to 35B parameter range. Iris-pro led or tied for the lead on all four benchmarks among systems in the roughly 400B parameter range.

Three Challenges Facing Search Agents

For conventional language models, the task, context, and compute budget are usually fixed before inference begins. A Search Agent, by contrast, has to choose its own search terms, read and filter information, judge whether the evidence is sufficient, and keep adjusting its reasoning path based on what it finds.

That brings Search Agents closer to real-world tasks, but it also creates difficulties around data, training, and evaluation.

The first is data construction. An ideal training question has to satisfy two conditions: the model cannot answer it directly without external material, but can arrive at a single answer once it has the relevant evidence. At the same time, the question cannot be solvable through simple keyword matching.

The second is training complexity. Long-horizon search tasks often involve many tool calls and reasoning steps, and a handful of extremely long trajectories can slow down synchronous training. If scoring and content summarization rely on external APIs, network volatility also undermines training stability.

The third is the lack of a consistent evaluation standard. A Search Agent's final score depends not only on the model itself but also on the tools, context management, number of search rounds, and the judge model. When test conditions differ, it is hard to tell whether a gap in scores comes from model capability or from the surrounding system design.

Iris's training approach is built around these three challenges.

Building Training Data Backwards From Web Links

Rather than simply collecting existing search questions, Iris generates training tasks in reverse from the hyperlink structure between web pages.

The system starts with a seed page, expands outward through its external links to related pages, and builds a local web graph. It then extracts entities and their relationships from that graph and generates questions that require multi-step reasoning across several pages.

To stop the model from simply searching for entity names given in the question, the system also applies "reference abstraction" to the entities involved. Apart from the final answer, every other entity in the question is rewritten as a descriptive phrase, preventing the model from finding the answer through string matching.

For example, instead of naming an institution or person outright, the question describes it by founding date, research focus, or partnerships. The model has to interpret these clues first, then work its way toward the target entity.

Once a question is generated, it goes through two rounds of validation.

The first is a difficulty check: if a reference model cannot answer the question without external material, the question is considered hard enough.

The second is a solvability check: if the reference model produces a correct and unique answer once given the relevant pages and evidence, the question is verifiable.

Only questions that are both hard to answer closed-book and solvable once evidence is provided are kept as training data. The entire pipeline runs automatically through models rather than manual question writing, which leaves room for further scaling.

Alternating SFT and RL Training

During training, Iris alternates between supervised fine-tuning (SFT) and reinforcement learning (RL), an approach the team calls "SFT-RL Climbing."

In the supervised fine-tuning stage, the team first has a stronger teacher model complete tasks using real search tools, producing full search trajectories. Those trajectories then pass through several layers of filtering.

Trajectory-level filtering checks whether the task was completed, whether the answer was correct, whether there were repeated searches or abnormal outputs, and whether search depth was sufficient. Trajectories with duplicate calls, useless loops, garbled text, or shallow searching are discarded.

At the turn level, the system examines whether each individual search and reasoning action was reasonable. Redundant searches, incorrect tool calls, and reasoning steps inconsistent with the action actually taken are masked out.

In the reinforcement learning stage, Iris trains directly in a live search environment. To limit the impact of unstable external services, the team deployed reward judging and retrieved-content summarization inside the training cluster rather than relying on external APIs.

Long-horizon search also produces a small number of extremely long tasks. The usual approach is to discard unfinished trajectories outright; Iris instead interrupts the task at the request level and resumes from the already-submitted prefix in the next training round, wasting less of the computation already done.

After each reinforcement learning round, the team feeds trajectories that were difficult but ultimately successful, along with those that were efficient and used fewer search steps, back into the next round of supervised fine-tuning. As the model improves, training tasks shift toward harder problems, producing a self-adapting training process.

Evaluation Under Uniform Conditions

Iris was tested on four benchmarks: BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE.

BrowseComp mainly tests a model's ability to locate long-tail entities through multiple indirect clues and produce an answer. BrowseComp-ZH focuses on Chinese-language web search. DeepSearchQA assesses how well an answer covers the relevant evidence. HLE tests expert-level reasoning across disciplines; Iris was evaluated on the text-only portion.

To reduce the influence of differences between inference frameworks, the team standardized the tool set, context length limit, and judging method across tests, and used a single ReAct agent with no multi-agent collaboration and no test-time self-verification.

Under the default "discard-all" context management setting published in the paper, Iris-mini scored as follows:

  • BrowseComp: 82.2
  • BrowseComp-ZH: 84.8
  • DeepSearchQA: 86.9
  • HLE: 52.3

Its BrowseComp, BrowseComp-ZH, and HLE scores were the highest among comparison systems in the 30B to 35B parameter range. Its DeepSearchQA score of 86.9 fell short of the 89.5 posted by a comparable system.

Iris-pro scored as follows:

  • BrowseComp: 88.6
  • BrowseComp-ZH: 85.1
  • DeepSearchQA: 92.9
  • HLE: 56.4

Its BrowseComp, DeepSearchQA, and HLE scores were the highest in its class, while its BrowseComp-ZH score tied for first with a comparable system.

The paper also notes that Iris-mini's BrowseComp performance approaches that of some models at the 1T parameter scale. That conclusion applies only to the specific test conditions published in the paper, and should not be read as the model matching those larger systems across all tasks.

How Context Management Affects Search Performance

During long-horizon search, a model can run out of context space before finishing the task. Many Search Agents therefore use context management mechanisms to compress, clear, or reorganize material they have already searched.

Iris was tested both without context management and with a "discard-all" strategy, which clears accumulated tool history once context hits a threshold and restarts the search from the original question.

With context management enabled, both Iris-mini and Iris-pro scored higher. The gain was larger for the smaller model, since smaller models typically need more interaction steps to finish the same task and hit context limits more easily.

The team argues that a Search Agent's final performance should be assessed with both the model and the inference framework in view. A search score published on its own may not fully reflect the model's own capability.

Search Skills Transfer to Other Agent Tasks

Beyond search benchmarks, the AllSpark team observed positive transfer from Iris's training data and search-specialized models to tasks they were never specifically trained on.

These include general tool-calling tests and Cowork-style office collaboration tasks, covering benchmarks such as BFCL, τ-bench, OfficeQA, and APEX.

The team argues that search training teaches more than web browsing: it also builds general capabilities such as acting under incomplete information, continuously gathering evidence, and revising plans based on feedback. Search, on this view, may not be an isolated vertical skill but a foundational one that carries over to other agent tasks.

What Comes Next

The AllSpark team says the Iris project will continue open-sourcing model weights along with its data construction, training, and evaluation pipelines, so researchers can reproduce and build on the work.

For now, the models and evaluation code for Iris-mini and Iris-pro are public, while the full data construction and training pipeline is still being prepared.

For Search Agents, model size is only one factor in capability. High-quality data, long-horizon training, context management, and consistent evaluation all shape the final result. Iris's experiments suggest that with more targeted training methods, mid-sized models have a shot at approaching much larger ones on specific search tasks.