From Paper to Program: Overcoming the AI Coding Bottleneck in Science

by Prof. Yi Zhou(周毅) (Institute of Physics, Chinese Academy of Sciences)

Wednesday 10 Jun 2026, 10:00 → 11:00 Asia/Shanghai

Tsung-Dao Lee Institute/N6F-N600 - Lecture Room (Tsung-Dao Lee Institute)

Host: Prof. Xin Liu  

Venue: TDLI Meeting Room N600

Tencent Meeting link: https://meeting.tencent.com/dm/DrmLEwwfkBNr 

Meeting ID: 239737182, no password

 

Abstract:

Large Language Models (LLMs) are highly capable at general programming, yet they consistently fail to translate complex scientific papers into working code. In this talk, I will demonstrate that the true bottleneck is not the AI's coding ability, but knowledge externalization. Research papers are written for human experts and routinely omit the tacit, computational "how-to" knowledge required to actually build the software.

To solve this, we designed a multi-stage "Virtual Research Group" workflow. Instead of asking the AI to code directly from a paper, an intermediate AI agent first drafts a rigorous technical blueprint. Tested on a notoriously difficult quantum physics algorithm (DMRG), this approach boosted the AI's success rate from 46% to 100% across 16 different model combinations—compressing weeks of graduate-level coding into under 24 hours. 

 

Biography:

Yi Zhou is a professor at Institute of Physics, Chinese Academy of Sciences. He received his B.S. degree and Ph. D. in physics from Tsinghua University in 1998 and 2004 respectively. After postdoctoral journey, he became a faculty member in Zhejiang University in 2009 and moved to current position in 2019. His main research interests include quantum many-particle physics and theoretical condensed matter physics.