Iris search agents released at two sizes with full training recipe

Researchers have released Iris-mini and Iris-pro, two open-source search agents built to answer questions that require hunting down information across the web rather than reciting memorized facts. Iris-mini has 35 billion total parameters with 3 billion active for any given query, and Iris-pro has 397 billion total parameters with 17 billion active, a design known as mixture-of-experts that switches on only part of the network per task. The paper, titled “Iris: Climbing to the Search Frontier,” describes both models and the full pipeline used to build them.

The training data was built backward from the structure of a web corpus. The authors say they mapped hyperlink chains outward from a seed page into an “entity graph,” then wrote multi-hop questions that require following several of those links to reach an answer. Every entity in a question except the final answer was rewritten into a descriptive reference, so a model could not shortcut the puzzle by matching text strings. A question only made the final set if a reference model failed to answer it from memory alone but succeeded once given the supporting evidence, a check meant to confirm the question truly required search rather than recall. Those questions were converted into full search trajectories, then filtered at both the whole-trajectory level and the individual-turn level before being used for supervised fine-tuning.

After that initial supervised pass, the models were refined with reinforcement learning against a live search environment, with a reward judge and a summarizer for search results run inside the training cluster itself. Rollouts that ran too long were paused at the request level and resumed from where they left off at the next training step, rather than discarded outright. The team alternated supervised fine-tuning and reinforcement learning in a cycle they call “SFT-RL climbing,” feeding the hardest problems each reinforcement-learning round solved, plus its most efficient solutions, back into the next supervised round.

Scores rise sharply with context management

The paper singles out one factor it says accounts for more of the score gap between systems than most of the differences reported between rival models: how an agent manages its context window, the working memory it has available, during a task. To isolate that effect, the authors tested every benchmark twice, once with inference-time context management turned on and once off, while holding the same tools, context limit and judge across both runs. With management enabled, Iris-pro scored 88.6 on BrowseComp, 85.1 on BrowseComp-ZH, 92.9 on DeepSearchQA and 56.4 on HLE, a wide-ranging knowledge and reasoning test called Humanity’s Last Exam. Iris-mini scored 82.2, 84.8, 86.9 and 52.3 on the same four tests. The paper describes these as the strongest results yet reported among open-source search agents at each model’s parameter scale. All scores came from a single ReAct-style agent, a method that alternates reasoning steps with actions like web searches, working alone without sub-agents or extra verification steps at test time.

The authors say they plan to release the model weights along with the complete recipe: the data construction method, the training procedure and the evaluation setup. That release had not yet appeared alongside the paper, so how the models hold up on tasks outside the paper’s own benchmarks remains to be seen.