Beyond the benchmark: How an adaptive AI approach drives scientific discovery

Microsoft’s Discovery Engine with CLIO achieved higher scores than other agentic harnesses on Agent’s Last Exam, a rigorous evaluation of long-running, tool-using professional tasks. The system scored 61.6% in health and medicine, 75.2% in physical sciences, and 64.6% in life sciences, reflecting its ability to pursue multiple hypotheses and adapt strategies based on evidence. CLIO enables independent reasoning paths, comparison of findings, and resolution into a single evidence-backed result, determining when to explore further, change strategy, or involve domain experts.

The platform is designed for organizations tackling problems without predefined workflows or known answers, such as materials teams balancing performance, safety, and cost. It supports researchers navigating incomplete evidence, competing objectives, and specialized tools while preserving traceability and governance standards already in use. Microsoft Discovery combines scientific hypothesis testing with engineering rigor, allowing multiple reasoning paths and a diverse model ecosystem to enhance adaptability.

Discovery Engine with CLIO has already supported the discovery of a novel organic redox flow battery, demonstrating real-world impact beyond benchmark results. The approach is applicable to design simulation, formulation optimization, materials discovery, and lab automation, helping organizations shorten research cycles without sacrificing rigor. The system is built to integrate with existing tools, data, and review processes, enabling researchers to accelerate innovation while maintaining transparency and expert validation.

Microsoft Discovery aims to bring agentic discovery to R&D organizations across industries and the scientific community, emphasizing iterative, collaborative, and adaptive problem-solving. While the platform is still early in development, this benchmark milestone highlights the potential of AI designed for the realities of scientific discovery. The company invites researchers and partners to explore how adaptive AI can expand their capabilities in tackling complex challenges.