Xueersi Working on a Large Math Model Called MathGPT

Image source: Visual China

BEIJING, May 6 (TMTPOST) — The popularity of ChatGPT has spilled over to the education and training industry in China.

On Friday, TiPost App learned from Xueersi that the company is developing a self-developed mathematics model named MathGPT, which is aimed at global mathematics enthusiasts and research institutions. The model is based on solving and teaching algorithms in the mathematics field, and has achieved phased results. The company plans to launch product-level applications based on the self-developed model later this year.

Xueersi has made MathGPT the core project of the company, which is led by CTO Tian Mi. Before the Spring Festival this year, the project had already launched corresponding team building, data and computing power preparation, and technical research and development.

In addition, Xueersi has also launched the building of a team in Silicon Valley, the United States, and will establish an overseas algorithm and engineering team to recruit outstanding AI experts globally.

The head of Xueersi’s AI team believes that the decision to build a self-developed model team was made because OpenAI, a US company, released the large language model GPT-4 in March this year, and domestic companies Baidu and Alibaba also released their own large model products. However, the general language model is more like a “liberal arts student” and performs well in tasks such as language translation, summarization, understanding, and generation, but has obvious shortcomings in solving, teaching, answering, and recommending math problems.

They found that these large models often make mistakes in solving math problems, and some math problems can be solved, but the methods are more suitable for adults and cannot be adapted to the knowledge structure and cognitive level of age-appropriate children.

“These shortcomings are determined by the characteristics of the LLM model,” said the head of Xueersi’s AI team. The LLM model comes from training on massive amounts of language text, so it is best at language processing. The industry tends to use LLM models for reading and writing applications, but if you want to make breakthroughs in math ability, you need to develop new large models.

Therefore, Xueersi decided to form a team to specialize in MathGPT, a large model in the mathematics field, and use its many years of accumulation in mathematics and AI to do the foundational work for mathematics in the era of AI large models, aimed at mathematics enthusiasts and research institutions worldwide.

Xueersi hopes to address and overcome three problems with MathGPT in large language models: firstly, the need to solve problems correctly, as GPT results often contain errors; secondly, the need for stable and clear problem-solving steps, as GPT’s problem-solving steps are different each time and generated content is often redundant; and thirdly, the need for interesting and personalized explanations, as GPT’s explanations are too “academic” and mechanical, which is not user-friendly for children’s learning experience.

As an education and training company that has been early in the field of AI, TiPost App has learned that in 2017, Xueersi established an AI lab for artificial intelligence research and development.

Starting from math extracurricular tutoring, Xueersi has accumulated 20 years of experience in math teaching and a vast amount of math-related data, which are essential for training MathGPT.

In addition, Xueersi believes that its overseas business, Think Academy, is popular among math enthusiasts in several countries and regions worldwide, and its students have performed well in international math competitions such as IMO and AMC each year, with several students winning gold medals in the International Mathematical Olympiad.

According to TiPost Focus, Xueersi’s learning machine will launch an “AI Assistant” soon, which includes related functions such as a composition assistant, a spoken language assistant, a reading assistant, and a math assistant. The AI product will start internal testing on May 11.

In the field of education, platforms such as Duolingo, Quizlet, and Khan Academy have embraced large models and collaborated with OpenAI to fine-tune and interface with GPT large models to enhance their existing product experiences.

In the field of math, products such as Photomath, Microsoft Math, Mathway, and WolframAlpha, which focus on math calculations, mainly utilize traditional AI technologies and databases to solve math problems. Companies that take the AGI route attempt to make general LLM “more knowledgeable in math,” such as GPT4, which performs better in math tasks than the previous version, and Google’s Minerva model, which is specifically optimized for math problems.

In addition to Xueersi, Netease Youdao plans to deploy large models in the education scene in China. Its ChatGPT model, developed based on “Ziyue,” will also be launched soon, and its AI spoken language teacher and Chinese composition correction DEMO, based on “Ziyue,” have been completed and will soon be open for internal testing.

On the same evening as Xueersi’s announcement, Netease Youdao also released a spoiler video for its AI spoken language teacher developed based on the “Ziyue” large model.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App

​Read More