Upstage to develop math-specific LLM with Qanda

South Korean artificial intelligence tech startup Upstage will challenge global math solver AIs such as OpenAI’s ChatGPT and Google’s Photomath with its own math-specific large language model (LLM), which would be the first of its kind in the country.  

Upstage announced on Wednesday it has agreed with Masspresso, the operator of Korean AI-powered learning platform Qanda, to develop a math domain-specific private LLM.

This is part of the two AI startups’ partnership with Korean telecom leader KT Corp., which in September invested 10 billion won ($7.7 million) in each of the AI startups to bolster its hyperscale AI capabilities.

KOREA’S FIRST MATH-SPECIFIC LLM

Domain-specific LLMs are designed to perform clearly defined tasks by conveying the most significant features of a certain industry through an understanding of its unique context, jargon and intricacies.

Compared with a generic language model that is relatively weak at understanding specific terminology, these specialized models can enhance user experiences and unlock unprecedented value for enterprises across different sectors. An increasing number of AI companies are rushing to develop domain-specific LLMs in response to growing demand.  

Under the agreement, Upstage will develop and fine-tune the math-specific LLM while Qanda will provide mathematics data on which Upstage’s AI will train.

Upstage said it will develop a math-specific LLM that will outperform not only generic LLMs with less hallucination but also other math-solver AIs such as OpenAI’s ChatGPT, Google’s Photomath and an Israeli startup's Symbolab in solving math problems.

The Korean startup said the country’s first math-solver LLM trained on Qanda’s large-scale math data will be able to understand complex mathematical formulas and concepts that are difficult to explain in natural language.

In July, Upstage’s language model, training with Meta’s LLaMA 2, topped the HuggingFace Open LLM Leaderboard, a standard for open-source LLM evaluation of over 500 LLMs, after beating OpenAI’s ChatGPT powered with GPT-3.5.

AI-powered Qanda allows its users to get answers and solutions from the photos of math or science problems uploaded by users. The app is popular not only in Korea but also in other countries such as Vietnam. 

PRIVATE LLM EXPANSION  

With a math-specific LLM, which would also be the Korean startup’s first domain-specific private LLM in the education sector, Upstage expects to accelerate the expansion of its private LLM applications across other industries.

A private LLM boasts a strong data protection and security system as it is specifically trained on a company's private internal dataset.

It is highly sought after because it is said to have less risk of AI hallucination, which refers to generation of misleading information that's presented as accurate.

With its own LLM Solace, Upstage has joined online e-commerce ConnectWave, Lotte Shopping Co. and Metaverse Entertainment in developing private LLMs that can be applied to their respective businesses.

The Korean AI startup also plans to develop domain-specific private LLMs for the financial and manufacturing sectors.

Write to See-Eun Lee at see@hankyung.com
Sookyung Seo edited this article.