arxiv.org/abs/2508.20729 Re4: Scientific Computing Agent with Rewriting, Resolution, Review and Revision
The paper presents a three-LLM agent (Consultant, Reviewer, Programmer) that rewrites problems, writes code, and iteratively debugs it for scientific computing tasks like PDEs and flow past a cylinder. The review loop raised execution success of DeepSeek R1 from 59% to 82%, ChatGPT 4.1-mini from 66% to 87%, and Gemini-2.5 from 60% to 84%.
— via arXiv — Physics, Ao Cheng, Lei Zhang, Guowei He
