AllSpark's Iris-mini, Iris-pro lead open-weight search agents

AllSpark's release stands out not just for benchmark scores but for how the training data was built. Rather than collecting questions from existing datasets, the team reverse-engineered multi-step questions from the link structure of web pages, paraphrasing every term except the final answer so that no clue can be resolved through a simple text search. A reference model then filtered for questions that required tools to solve, keeping the tasks hard and verifiable.

The paper also makes a pointed argument about evaluation. The team found that runtime context management on common benchmarks often makes a bigger difference than the reported gaps between systems. For Iris-mini, context management boosted BrowseComp scores by up to 21.2 points, not because of a smaller token budget but because the smaller model consumes context faster and hits the limit more often. This suggests that reported benchmark gaps between models may partly reflect differences in scaffolding rather than model quality alone.

Beyond search, the authors report that both the generated training data and the specialized models improved performance on tasks they were never trained for, including general tool use and office work. That points to search as a foundational skill rather than a narrow specialty. The main open question is whether the team's planned release of data construction and training pipelines will let others verify and build on these results.