Summary
ComputeGPT is an accurate chat model designed to solve math and science problems. It's known as a math chat model or a computational chat model, due to it not being a typical probabilistic LLM, but rather generating code on-demand to solve complex problems, from homework to textbook to general numerical problems. It's also known as MathGPT and ScienceGPT.
ComputeGPT is a computational chat model that accurately answers math and science problems. It outperforms GPT-4 with Internet (Bing), GPT-3.5 (ChatGPT), GPT-3, and Wolfram Alpha NLP, scoring 98% across a suite of numerical problems that Bing AI could only answer at 64% and Wolfram Alpha NLP at 56%.
![ComputeGPT Evaluation](https://sites.utexas.edu/uil/files/2023/05/image.png)
One of the focuses for this project was on accuracy, as current AI solutions have been shown to have inaccurate or misleading answers, known as “hallucinations”. Our model runs verifiably correct code on-demand to ensure that the answer it gives is completely accurate. This project fixes the “probabilistic accuracy” in language models created by LLM’s inherent architecture: a focus on generating the next word. By instead generating code and running it, ComputeGPT demonstrates state-of-the-art performance in homework, textbook, and all kinds of numerical problems, including mathematics, physics and medicine. ComputeGPT is a math chat model for computing any numerical answer.
See our paper here: https://arxiv.org/abs/2305.06223.
See our GitHub here: https://github.com/urbaninfolab/ComputeGPT.
Unlike other LLMs, ComputeGPT is available at no cost and logs no information about its users. Try it out now at https://computegpt.org.
Team Members:
Junfeng Jiao (Department of Architecture), Ryan Hardesty Lewis (Department of Mathematics)
(Click Here for More Project Information)
![Math chat model MathGPT ComputeGPT](https://sites.utexas.edu/uil/files/2023/05/hero-image.458e505b.png)