Iris-mini and Iris-pro: New AI Search Agents That Navigate the Web Like Humans
Summarized by AI from reporting by ArXiv cs.AI, published under our editorial policy.
Researchers introduce Iris-mini (35B parameters) and Iris-pro (397B parameters), two AI search agents trained to answer complex multi-hop questions by navigating web hyperlinks. Unlike keyword-based search, these models reason through entity graphs and outperform reference models on closed-book tasks.
Key takeaways
- Iris-mini (35B parameters, 3B active) and Iris-pro (397B parameters, 17B active) are AI search agents trained to navigate web hyperlinks via multi-hop reasoning.
- The training pipeline reverse-constructs tasks from web hyperlink structure, rewriting non-answer entities into descriptive references to prevent string-matching shortcuts.
- Both models outperform reference models on closed-book tasks, solving complex queries only when supporting evidence is provided.
Researchers have introduced Iris-mini and Iris-pro, two new AI search agents designed to revolutionize how we find information online. Unlike traditional search engines, these models are trained to understand and navigate the web's hyperlink structure, answering complex questions by following multi-step reasoning paths.
Training on Multi-Hop Entity Graphs
Iris-mini and Iris-pro are trained using a unique data pipeline that mimics how humans search for information. The process starts with a seed page and its out-links, creating an entity graph. From this graph, the models construct multi-hop chains of tasks, ensuring that no question can be answered by simple string matching. This forces the models to understand context and derive answers through reasoning, much like a human would.
Performance Benchmarks: 35B vs 397B Parameters
Iris-mini, with 35 billion parameters (3 billion active), and Iris-pro, with 397 billion parameters (17 billion active), have been tested against traditional search engines. The results show that these AI agents can answer questions that even advanced reference models struggle with when operating in a 'closed-book' mode (without direct access to the information). Once given the supporting evidence, the models excel at solving complex queries, demonstrating a significant leap in search technology.
Why It Matters for Everyday Users
For everyday users, Iris represents a shift from keyword-based searches to a more intuitive, human-like search experience. Imagine asking a question like, 'What are the environmental impacts of electric vehicles compared to gas cars?' and receiving a detailed, well-reasoned answer that considers multiple sources and perspectives. This could make research, decision-making, and even casual browsing more efficient and informative.
Current Status and Availability
While Iris is still in the research phase, you can stay updated on the latest developments in AI search by following arXiv's cs.AI section. This is where cutting-edge research is often first published, and it's a great resource for anyone interested in the future of AI.
Frequently asked
- Are Iris-mini and Iris-pro available for public use?
- Not yet. The models are still in the research phase and have not been released for public use.
- How do Iris models differ from traditional search engines?
- Traditional search engines rely on keyword matching, while Iris models understand context and follow multi-step reasoning paths to answer complex questions.
- What does '35B-A3B' and '397B-A17B' mean in the paper?
- These refer to the total parameter count and the active parameters per forward pass. Iris-mini has 35 billion total parameters with 3 billion active, and Iris-pro has 397 billion total with 17 billion active.