At the 2026 Inclusion·Bund Summit, held from September 9 to 12 at the Shanghai Huangpu Expo Park under the theme "Co-Creating the AI Economy," Mengdi Wang, Director of the Princeton University AI Innovation Center, shared her perspective that large language models have yet to deliver genuinely novel scientific discoveries in core disciplines like biology, chemistry, and physics, particularly within experimental fields.

Wang compared the pursuit of foundational scientific frontiers to searching for a star never before observed in the universe—a celestial body absent from any training dataset. She explained that from the standpoint of knowledge acquisition and fitting, LLMs capture the primary modes of probability distributions while neglecting the long-tail segments where true human innovation often occurs. As a result, current AI systems have not advanced to the scientific frontier and cannot yet break through existing probability distributions to locate that elusive "new star."

Wang emphasized that AI's full potential can only be realized when the digital world forms a closed loop encompassing the entire research process of the physical world. Only then can laboratory environments across disciplines become verifiable spaces, potentially extending the current training paradigm of LLMs to facilitate new scientific discoveries. She noted that while AI models learn established facts and patterns, scientific discovery involves seeking entirely new possibilities beyond probability distributions, requiring sustained exploration and investment at the infrastructure level to bridge the vast gap between known laws and unknown findings.