Microsoft Discovery is a platform designed to support research and development (R&D) organizations by leveraging agentic AI. The core principle is to move beyond single-answer models and embrace a dynamic approach to scientific discovery, mirroring the way researchers and engineers actually work. This involves pursuing multiple hypotheses, validating them against evidence, and adapting strategies based on new information.
The system’s performance was evaluated on Agent’s Last Exam, a demanding benchmark for long-running, tool-using professional tasks. Microsoft Discovery Engine with CLIO achieved scores of 61.6% in health and medicine, 75.2% in physical sciences, and 64.6% in life sciences, surpassing other agentic harnesses. This success is attributed to CLIO, a Cognitive Loop via In-Situ Optimization system.
CLIO enables independent reasoning paths, allowing the system to compare and share learning, and resolve the strongest trajectory into a single, evidence-backed result. Crucially, CLIO can determine when to continue exploring, change strategy, utilize a different model, or involve a domain expert. This adaptive reasoning loop is designed to handle the complexities of real-world scientific challenges, including incomplete evidence and competing objectives.
Beyond the benchmark results, Microsoft Discovery is being applied to various research areas. For example, the engine supported the discovery of a novel organic redox flow battery. The approach has potential applications in design simulation, formulation optimization, materials discovery, and lab automation – areas where organizations seek to accelerate research cycles without compromising rigor. The platform is designed to integrate with existing tools, data, governance, and review processes, providing a systematic and transparent pathway to outcomes.
Microsoft Discovery aims to redefine R&D by combining the scientific mindset with engineering rigor. The adaptive reasoning capabilities of CLIO, combined with the platform’s enterprise-grade design, offer a powerful tool for organizations across all industries and the scientific community. The initial benchmark milestone demonstrates the potential of AI built for the way discovery actually happens: iteratively, collaboratively, and adaptively.
