Upstage-Kwanda-KT Surpass ChatGPT with Math-Specialized Model

Jointly Developed MathGPT (Tentative Name) Sets New World Record in Leading Math Ability Benchmark Test

Upstage announced on the 8th that ‘MathGPT (tentative name)’, a math domain-specialized AI model jointly developed with Masspresso, which operates the AI-based learning platform ‘Qanda’, and KT, has set a new world record by surpassing OpenAI’s ChatGPT and Microsoft (MS) models.

In November last year, Upstage and Qanda began developing MathGPT as part of a strategic partnership with KT. Upstage fine-tuned the natural language-based language model by training it on Qanda’s high-quality specialized math data, enabling it to solve complex math problems through logical reasoning and programming.

The two companies developed MathGPT and achieved impressive results that surpassed MS’s ‘ToRA 13B’, the strongest model in its class, on representative benchmark tests evaluating language models’ math abilities such as ‘MATH’ and ‘GSM8K’. MathGPT simultaneously achieved top performance on the MATH benchmark, which consists of 12,500 challenging math competition problems, and the GSM8K benchmark, which tests arithmetic operations with 8,500 elementary school math problems.

MathGPT surpassed ChatGPT’s performance on average across benchmark tests and even outperformed GPT-4 on the MATH benchmark. In the high-difficulty math domain, a domestically developed small-sized model outperformed big tech companies like OpenAI and MS.

Having confirmed achievements in the education sector through MathGPT, Upstage plans to lead the reorganization of the large language model (LLM) market with its own model ‘Sola’. Covering various industries including finance, distribution, healthcare, and entertainment, the company will focus on strengthening its global presence as a stepping stone for full-scale overseas expansion beyond Korea.

Kim Seong-hoon, CEO of Upstage, said, “It is meaningful to have developed the world’s best math-specialized language model surpassing ChatGPT through collaboration with Qanda and KT,” adding, “Going forward, Upstage will lead generative AI innovation in various fields based on its global No.1 LLM technology.”