Iris: A New AI Paradigm for Autonomous Machine Learning Engineering
A recent paper published on arXiv (2608.02143) presents Iris, a framework designed for autonomous machine learning engineering through an inquiry-revision loop. The researchers contend that current LLM-based agents rely on solution-focused searches, structuring candidate-solution enhancements via trees, graphs, or chains, which influences how information is gathered and processed. They suggest a new information framework where a dynamic information state reflects the system's comprehension of the task, directing solution enhancements. Iris formulates localized action plans based on the existing information state and employs epistemic actions to explore critical unknowns without altering the existing solution. For managing information, it integrates observations from various experiments into task knowledge made up of revisable claims, aiming to enhance long-term autonomous research endeavors with constrained resources.
Key facts
- Paper arXiv:2608.02143 introduces Iris, an inquiry-revision loop for autonomous ML engineering.
- Iris uses an information paradigm instead of solution-centric search.
- Existing LLM-based agents use tree, graph, or chain structures for candidate-solution improvement.
- Iris generates local action plans from the current information state.
- Iris uses epistemic actions to probe decision-critical unknowns without modifying the retained solution.
- Iris synthesizes observations across experiments into task knowledge composed of revisable claims.
- The paper addresses long-horizon autonomous research tasks under limited budgets.
- The paper is published on arXiv.
Entities
Institutions
- arXiv