This paper introduces DeepScientist, an autonomous AI system designed to conduct goal-oriented, long-term scientific discovery by modeling the process as a Bayesian Optimization problem. This system operates through a hierarchical cycle of hypothesizing, verifying, and analyzing, leveraging a cumulative Findings Memory to balance exploration and exploitation efficiently. By consuming over 20,000 GPU hours, DeepScientist generated thousands of ideas and successfully validated methods that progressively surpassed human State-of-the-Art (SOTA) methods across three frontier AI tasks, demonstrating a rapid evolution curve compared to human research. The paper provides evidence that an AI can achieve SOTA-surpassing scientific breakthroughs, although the analysis reveals that effective validation and filtering of generated ideas remains the primary bottleneck due to the low success rate of novel hypotheses. The authors emphasize that the future of this work lies in improving discovery efficiency and fostering human-AI collaboration for strategic guidance.
References:
