ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #Iris #Open-Source Models #Search Agents #Multi-Hop Reasoning #Reinforcement Learning

Iris Team Releases Iris-mini and Iris-pro: Open-Source Search Agents with SOTA Performance

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 6 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:The Iris Research Team has unveiled Iris-mini and Iris-pro, two open-source search agents trained at scales of 35B-A3B and 397B-A17B, respectively. These agents leverage multi-hop reasoning and sophisticated question generation techniques, achieving state-of-the-art results on benchmarks like BrowseComp, DeepSearchQA, and HLE. The team plans to release the model weights along with the complete data construction, training, and evaluation recipes, providing AI developers with powerful tools for bu


Technical Highlights

  1. Model Architecture and Scale

    • Iris-mini: 35B parameters, A3B architecture
    • Iris-pro: 397B parameters, A17B architecture These models are designed to handle complex multi-hop reasoning tasks with their large scale and advanced architecture.
  2. Training Methodology

    • Reverse-constructed tasks from hyperlink structure: Utilizing the hyperlink structure of web corpora to generate multi-hop reasoning chains and distilling entity graphs for training data.
    • Descriptive rewriting: All non-answer entities are rewritten into descriptive references to prevent direct resolution through string matching.
    • SFT-RL Climbing: Alternating between Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to optimize the policy against live search.
  3. Evaluation and Results

    • Iris-mini and Iris-pro achieved strong results on benchmarks like BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE, with scores of 82.2/84.8/86.9/52.3 and 88.6/85.1/92.9/56.4, respectively.
    • The team emphasized the importance of inference-time context management, which significantly boosts performance.
  4. Open-Source Plans

    • The Iris team intends to release the model weights along with the complete data construction, training, and evaluation recipes, providing the AI community with resources for reproducible research.

Industry Impact and Developer Recommendations

  • Impact on AI Search: The release of Iris-mini and Iris-pro marks a significant advancement in open-source search agents, offering new possibilities for building smarter search engines and knowledge management systems.
  • Implications for AI Developers: Developers can leverage the open-source resources from the Iris team to rapidly build and customize search agents, particularly for applications requiring complex query handling and multi-hop reasoning.
  • Future Research Directions: The team suggests further optimizing inference-time context management, exploring more efficient multi-hop reasoning algorithms, and expanding the models' applications across different languages and domains.

Conclusion

The release of Iris-mini and Iris-pro not only demonstrates the immense potential of AI in the search domain but also provides the open-source community with powerful tools and resources, driving the adoption and application of AI technologies.


Source: Hugging Face Daily Papers (2026-09-03)

— END —

Tags: #Iris #Open-Source Models #Search Agents #Multi-Hop Reasoning #Reinforcement Learning

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…