Beyond the benchmark: How an adaptive approach drives scientific discovery

For research and development (R&D) organisations, the potential of agentic AI represents more than just a simple solution. It’s a revolutionary way to tackle intricate scientific and engineering challenges. This approach includes exploring various hypotheses, validating them through evidence, learning from failures, and adapting as new insights emerge.

The unique aspects of this agentic discovery process have become a primary focus of Microsoft’s research, forming a key design principle for Microsoft Discovery—our platform for organisations committed to cutting-edge R&D.

Evaluating Adaptive AI for Scientific Discovery

Recent benchmark results illustrate how this vision is turning into reality. On the Agent’s Last Exam, which rigorously assesses the performance of long-term, tool-oriented tasks, the Microsoft Discovery Engine with CLIO (Cognitive Loop via In-Situ Optimization) outperformed other evaluated agentic systems across three scientific fields: registering 61.6% in health and medicine, 75.2% in physical sciences, and 64.6% in life sciences.

This achievement reflects Microsoft’s foundational research into what distinguishes agentic discovery. CLIO facilitates independent reasoning paths that allow exploration of problems, exchange of insights, and converging on a robust, evidence-based conclusion. It can identify when to continue investigating, alter the strategy, switch models, or involve a domain expert.

The CLIO benchmark blog post goes into detail about this adaptive reasoning method. More broadly, this fundamental innovation for scientific discovery, driven by agentic AI, is accessible to R&D organisations in all industries and to the scientific community through Microsoft Discovery, not just as a theoretical development but as a vital tool for practical R&D efforts.

The Importance of Adaptive Reasoning in Scientific Discovery

Many of the most challenging scientific and engineering problems lack a clear workflow or definitive answers. Researchers often face incomplete evidence, conflicting objectives, specific tools, and shifting constraints. For instance, a materials science team might have to juggle factors like performance, safety, costs, and manufacturability. In contrast, a life sciences team may need to synthesise existing literature, proprietary data, models, and experimental findings before deciding on their next validation step. Meanwhile, an engineering team could be tasked with exploring a vast design space without compromising on physical accuracy or traceability.

In these complex scenarios, relying on a single model output doesn’t suffice. Professionals require systems capable of reasoning over time, preserving evidence, questioning assumptions, and integrating seamlessly with the tools, data, governance, and review processes they already employ. Equally vital is the ability to understand how a conclusion was reached and where human judgment should intercede.

Microsoft Discovery was crafted as an enterprise-level platform for agentic R&D, seamlessly merging the scientific approach of hypothesis, experimentation, and refinement with the engineering discipline of problem breakdown, structured execution, and reproducibility. CLIO enhances this foundation by promoting a more adaptive reasoning cycle and a diverse range of model options, enabling researchers to explore multiple reasoning paths effectively.

Shifting from Benchmarks to Real-World Outcomes

The significant possibilities extend well beyond benchmark results into actual research settings. The Discovery Engine with CLIO has already facilitated the development of a new type of organic redox flow battery. This method holds immense promise across various domains, including design simulation (such as silicon chips), product formulation and process optimisation (for instance, within manufacturing and consumer packaged goods), materials and molecular discovery (which could spur sustainability initiatives and drug development), and lab automation—areas where organisations aim to reduce research timeframes while maintaining thoroughness and traceability.

Agentic discovery is not about replacing scientists and engineers. Instead, it enhances their ability to explore, accelerates their learning from evidence, and provides a more systematic and transparent pathway from concepts to outcomes that experts can scrutinise and validate.

To truly harness the vast potential of redefining R&D, we need a platform that integrates with the existing tools, data, governance, and review frameworks that researchers rely on. Microsoft Discovery meets this need by offering agentic discovery to R&D professionals across diverse industries and within the scientific community.

We are still at the beginning of this exciting journey, but this benchmark achievement highlights what is achievable when AI is designed to align with actual discovery processes: through iterative, collaborative, and adaptive methods. I eagerly anticipate the new discoveries that organisations, researchers, and partners will make moving forward.

FAQ

What is agentic AI? Agentic AI refers to artificial intelligence that can explore various hypotheses, adapt its strategies, and learn from evidence in a way that simulates human reasoning.

How does Microsoft Discovery support R&D? Microsoft Discovery provides a platform that integrates adaptive reasoning, allowing researchers to manage complex scientific problems more effectively.

Why is adaptive reasoning crucial in scientific discovery? Adaptive reasoning helps researchers navigate flexibility and changing conditions, ensuring they can address incomplete evidence and evolving objectives effectively.

What are some applications of agentic discovery? Agentic discovery can be applied in various fields, including materials science, drug discovery, product optimization, and lab automation, enhancing research efficiency and outcomes.