Recently, Westlake University launched an AI scientist named DeepScientist. This system completed the research achievements of human scientists over three years within just two weeks. DeepScientist not only autonomously generated more than 5000 scientific ideas and validated 1100 of them, but also successfully broke the latest human records in three cutting-edge AI tasks, demonstrating its strong research capabilities.
In the history of AI research, although many tools and systems have emerged, they are mostly auxiliary tools that cannot independently complete research work. Early AI systems such as PaperBench and Agent Laboratory mainly helped scientists reproduce papers or solve specific machine learning problems, while AlphaTensor optimized code performance through extensive trial and error. However, these tools often failed to question existing research paradigms.
With technological advancements, some fully automated AI scientist systems such as AI Scientist have emerged, proving that AI can complete the entire research cycle and discover new scientific achievements, but they still lack clear scientific goals and directions. In contrast, DeepScientist demonstrates target-oriented exploration and insight with its closed-loop and iterative process. The system first analyzes existing research methods, identifies their shortcomings, and then proposes novel and scientifically meaningful ideas.
The workflow of DeepScientist is designed as a three-stage cycle: first, the stage of generating new ideas, where the system extracts information from a large memory library and scores new ideas; next, the system uses the "upper confidence bound" algorithm to decide which idea to validate, selecting the highest-scoring idea for experimentation; finally, based on successful validation, the system writes detailed research reports, forming a closed loop.
In specific tasks, DeepScientist selected three cutting-edge AI research areas, including agent failure attribution, LLM reasoning acceleration, and AI text detection. In these three tasks, DeepScientist proposed new methods such as A2P, ACRA, and PA-Detect, not only surpassing existing SOTA (State Of The Art) records, but also demonstrating its excellent self-learning and innovative capabilities.
This groundbreaking research result further verifies the potential of AI scientists in the field of scientific research, and it may play an important role in broader scientific exploration in the future.
Project: https://github.com/ResearAI/DeepScientist
Key Points:
- 🚀 DeepScientist completed the research of human scientists over three years within two weeks, showcasing powerful research capabilities.
- 💡 The system can autonomously generate and validate scientific ideas through a closed-loop, iterative process, forming a complete research cycle.
- 🧠 DeepScientist successfully broke the latest human research records in multiple cutting-edge tasks, demonstrating the great potential of AI in the field of scientific research.
