Can AI agents conduct open-ended AI research? Early evidence from two case studies

Research from arXiv explores the capability of AI agents to conduct open-ended AI research. It introduces a novel evaluation method, 'shadow evaluations,' where AI agents address core research questions from high-quality, unpublished papers, with the original authors grading the output. This approach aims to provide clearer evidence on the potential for AI agents to automate AI research, contrasting with existing methods that are either too narrow or problematic.

Why it matters

The potential for AI agents to automate AI research has significant strategic implications for innovation cycles, resource allocation, and competitive advantage across various sectors. Understanding the capabilities and limitations of AI in open-ended research is crucial for shaping future research and development strategies, impacting how organizations invest in and leverage AI technologies.

Key insights

• Current methods for evaluating AI agents in research are limited, either focusing on narrow, verifiable tasks or relying on a flawed peer-review process.
• A new evaluation methodology, 'shadow evaluations,' is proposed, involving AI agents tackling central, open-ended research questions from unpublished papers.
• The output of AI agents in shadow evaluations is graded by the original authors of the papers.
• Two shadow evaluations were conducted using frontier AI agents over a six-day period on unpublished NeurIPS 2026 submissions.
• The study aims to measure progress towards the automation of AI research and the capacity for AI agents to undertake open-ended research tasks.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.27191