Dunamu, the operator of the virtual asset exchange Upbit, unveiled its global research achievements in the field of artificial intelligence (AI) for the year 2025 on December 18, 2025.
The Dunamu Machine Learning (ML) Team presented a research paper on text-to-SQL conversion at the main conference of the International Conference on Computational Linguistics 'COLING 2025,' held in Abu Dhabi, United Arab Emirates, in January. COLING is a globally renowned academic conference in the fields of natural language processing (NLP) and computational linguistics, known for its highly rigorous review standards.
The presented paper, "MCS-SQL: Leveraging Multiple Prompts and Multiple-Choice Selection For Text-to-SQL Generation," proposed a method to address the issue of response inconsistency in large language model (LLM)-based text-to-SQL conversion. By generating multiple SQL candidates using multiple prompts and then selecting the optimal result, the approach improved both accuracy and efficiency. This model demonstrated its technical excellence by ranking first on the global BIRD-SQL leaderboard, a benchmark for text-to-SQL conversion.
In addition, the Dunamu ML Team also presented its research achievements at 'SIGIR 2025 (The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval),' the most prestigious international conference in the field of information retrieval, held in Padua, Italy, in July. SIGIR is known for its intense competition, with only about 27% of submitted papers accepted this year.
The paper presented at SIGIR, "LLM as User Simulator: Towards Training News Recommender without Real User Interactions," proposed a method for generating virtual users using LLMs and training recommendation models without real user data. The 'LAUS (LLM As User Simulator)' framework developed by the Dunamu ML Team simultaneously addresses the burdens of large-scale user data collection and privacy concerns. Experimental results showed that LAUS outperformed zero-shot approaches for news recommendation without training data and also reduced response latency.
Dunamu has also expanded the application of AI technology to its actual services beyond research achievements. The Upbit Developer Center was completely revamped to introduce the AI-powered 'Upbit Assistant.' This service enhances developer convenience by learning from Developer Center documents to provide API examples, document summaries, and guide navigation. It also offers the llms.txt format to enable major AI tools such as GPT and Claude to utilize the latest Upbit API documentation, and has strengthened the data-driven development environment by adding the 'Upbit Historical Market Data' service, which allows users to download past market data.
Dunamu's investment information platform, Securities Plus, has also advanced its AI features. The recently added 'Newsroom' tab uses Dunamu's proprietary machine learning algorithms to comprehensively analyze the importance and market impact of news, prioritizing information essential for investment decisions. The 'Breaking' section quickly delivers key content through headlines and three-line summaries, while the 'Top News' section selects and provides the 'Top 5 Core News in 24 Hours' with the greatest impact during the day. Each news item displays related stock names and price change rates, enhancing the utility of information for investors.
Lee Dongjun, Head of the Dunamu ML Team, stated, "Based on the AI research achievements and service application experience accumulated this year, we will continue to strengthen our technological competitiveness next year to provide users with even more reliable services."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
