Microsoft is advancing the capabilities of artificial intelligence in scientific discovery with the development of CLIO, a self-adaptive reasoning system designed to be both controllable and explainable by domain experts. Unlike conventional large language models whose reasoning strategies are fixed during post-training, CLIO enables flexible, in-situ optimization: users can steer the model’s cognitive processes in real time, tailoring its approach to unique scientific challenges. By enabling reflection loops and dynamic reasoning, CLIO represents a paradigm shift away from static, post-trained behaviors toward interactive, user-driven discovery workflows.
Through evaluations on the demanding Humanity’s Last Exam (HLE) in biology and medicine, CLIO demonstrated notable performance improvements over state-of-the-art models. For example, on text-only HLE tasks, CLIO elevated OpenAI GPT-4.1’s accuracy from 8.55% to 22.37%, a 161.64% relative gain. This was achieved without requiring traditional reinforcement learning or additional post-training, instead capitalizing on runtime self-reflection and user-guided strategies. CLIO’s architectural innovations—such as enabling users to set thresholds for uncertainty and edit the agent’s beliefs—also are designed to address the dual challenges of model confidence and scientific rigor, making it possible to catch and correct errors instead of yielding only confidently wrong answers.
CLIO’s model-agnostic approach improves both explainability and reproducibility by documenting internal reasoning steps, uncertainty flags, and decision pathways. Scientists can adjust how much ‘thinking effort’ CLIO expends and direct the model’s reasoning style for different problems. Tests show CLIO can uplift even base models like GPT-4o to performance levels comparable with specialized reasoning agents, especially in biomedical domains. Microsoft envisions CLIO not just as a research breakthrough but as an enabler for trustworthy, user-driven artificial intelligence in broader fields—from drug discovery to finance and legal analysis. The innovation forms a cornerstone for the recently announced Microsoft Discovery platform, aiming to foster collaborative, transparent scientific advances in an era of increasingly capable artificial intelligence systems.
