LG AI Research announced that its self-developed artificial intelligence platform, “Exaone,” has demonstrated world-class performance by successively outperforming AI models from global Big Tech companies like Google and Alibaba in industrial data prediction. The achievement is seen as securing differentiated competitiveness with AI specialized for industrial sites—such as manufacturing, finance, and healthcare—amid a global AI landscape dominated by general-purpose large language models (LLMs).
On the 7th, LG AI Research revealed that “Exaone Tabular,” which analyzes and predicts data in table format, and “Exaone Forecast,” which predicts time-series data, each achieved state-of-the-art (SOTA) performance in global AI evaluations.
Exaone Tabular took the overall top spot in the categorical data prediction category of TabArena, a global leaderboard for structured data evaluation. It scored 1,760 ELO points, edging out Google’s latest model, “TabFM” (1,749 points), by 11 points. This model is a tabular foundation model (TFM) that directly learns the row and column structures of tables and the relationships between variables. Unlike existing general-purpose language models that convert tables into text for processing, Exaone Tabular can more efficiently analyze structured data from industrial sites, such as defect prediction and quality control in manufacturing processes, financial transaction records, and medical clinical information.
Notably, the model was trained on over 1 billion synthetic table datasets and has the advantage of not requiring full retraining when new data is introduced. Even when some data values are missing, it can predict those missing values by analyzing the relationships between surrounding data points.
The time-series forecasting model “Exaone Forecast” also stood out in global evaluations. LG AI Research stated that Exaone Forecast ranked first in the zero-shot category of “GIFT-Eval,” a time-series forecasting AI evaluation platform developed by Salesforce. Zero-shot evaluation measures how accurately a model predicts new data without additional training, and Exaone Forecast achieved the highest performance, surpassing models from Google and Alibaba. It also ranked second in the Agentic AI category, where the AI analyzes data and makes predictions autonomously.
Exaone Forecast is a time-series foundation model that predicts data changing over time, such as product demand, raw material prices, power consumption, and stock prices or exchange rates. It was trained on 25 billion real-world data points and 2 trillion synthetic time-series data points. A key feature is that it learned from virtual scenarios reflecting trends, seasonality, volatility, and structural changes to respond to shifts not present in historical data. It can be applied across various fields like manufacturing, finance, energy, and healthcare without needing to build separate models tailored to specific industries.
LG AI Research is already utilizing this technology within its affiliates, including LG Electronics and LG Energy Solution, for predicting product demand and raw material prices like lithium. Starting this year, in collaboration with Koscom and the London Stock Exchange Group (LSEG), it is also operating a stock market prediction service analyzing approximately 8,000 listed companies in South Korea and the United States.
LG AI Research plans to conduct proof-of-concept (PoC) projects in manufacturing, bio-healthcare, and finance during the second half of this year before moving into full-scale commercialization. Applications are slated to include battery cell defect detection, disease risk analysis, and the prediction of financial delinquencies, defaults, and abnormal transactions.
LG is expanding Exaone not as a single general-purpose model but as a family of industry-specific foundation models. Following Tabular for table data analysis and Forecast for time-series prediction, it unveiled “Exaone Path” last year, which analyzes pathology images. Medical AI and robotics foundation models are also currently under development, broadening the application scope to include technology that controls the physical tasks of robots.
“The focus of global Big Tech is shifting from general-purpose language models to specialized AI that solves problems on industrial sites,” said Lim Woo-hyung, co-director of LG AI Research. “We have proven that the data accumulated in industrial fields can be a differentiated competitive advantage for Korean AI.”
This achievement underscores the growing importance of AI models specialized in analyzing and predicting structured and time-series data generated in real-world industrial settings, distinct from general-purpose LLMs like OpenAI’s GPT or Google’s Gemini. LG AI Research plans to secure leadership in the industrial-specific AI market by expanding Exaone’s application range from industrial data analysis and prediction to medical diagnosis and robot control.