It beat Microsoft’s 'ToRA 13B', and even outperformed OpenAI's GPT-4 in the MATH benchmark

A screengrab shows the interface of Google-backed MathGPT application. — MathGPT

Mathpresso, the creator of Asia's leading AI-driven learning platform, QANDA, has proudly announced that its large language model, MathGPT, has set a new world record in mathematics, outperforming models from OpenAI and Microsoft.

MathGPT has secured the top position in benchmark evaluations assessing mathematical proficiency, such as the 'MATH' benchmark (12,500 challenging math problems) and 'GSM8K' benchmark (8,500 elementary school math problems). 

Notably, it surpassed Microsoft’s 'ToRA 13B', the previous record holder, and even outperformed OpenAI's GPT-4 in the MATH benchmark.

This achievement marks a collaborative effort between Qanda, Upstage, and KT, initiated in November of the previous year. Qanda contributed essential learning data to Upstage, facilitating the development of MathGPT. Upstage, in turn, employed its specialised solution to prevent hallucinations and fine-tuned the language model for logical inference.

MathGPT's success stands in contrast to models like ChatGPT, which, due to its training on extensive text data, often generates responses with inaccuracies, a phenomenon known as hallucination. 

In educational contexts, precision and reliability are crucial, and MathGPT's superior accuracy, particularly in mathematical domains, positions it as a breakthrough in AI-driven education.

The statement from Qanda expresses a commitment to further enhance MathGPT's accuracy and performance. The ultimate goal is to integrate it with their learning interface to introduce an 'AI Tutor,' serving as an assistant teacher. The plan is to extend AI tutor services to educational sites globally, revolutionising the education market.

Kim Seong-hoon, CEO of Upstage, emphasised the significance of developing a world-class, math-specific language model through collaboration. 

He stated, "We will lead generative AI innovation," highlighting their commitment to advancing the field of artificial intelligence. With major support from entities like Google, TikTok, and Softbank Ventures Asia, Qanda continues to play a significant role in reshaping education through innovative AI solutions.

Source link