LG AI Research said on Thursday that its in-house model for predicting tabular data, EXAONE Tabular, and its model for forecasting future data over time, EXAONE Forecast, each achieved state-of-the-art performance on global leaderboards.

Woo-hyung Lim (임우형), co-head of research at LG AI Research, said global big tech attention is rapidly shifting from general-purpose language models to industry-specific AI foundation models that solve real problems in industrial settings. He said the result proves that data accumulated across industrial sites, including manufacturing, can become a differentiated competitive strength for South Korean AI.

LG AI Research said EXAONE Tabular ranked first overall in the categorical data prediction category on TabArena, a leading global leaderboard that evaluates tabular data prediction used across industries.

EXAONE Tabular is a Tabular Foundation Model trained at scale on synthetic tabular data. Unlike general-purpose language models that convert tables into text for analysis, it can directly understand the structure of tables, made up of rows and columns, and relationships among data.

EXAONE Tabular scored 1,760 ELO points, beating Google’s TabFM, the latest model released last month, which scored 1,749.

Using EXAONE Tabular eliminates the need to retrain the entire model each time new data is provided. LG AI Research said it can quickly adapt to new problems with only a small amount of data and produce prediction results, making it highly useful even in industrial settings where data is insufficient.

LG AI Research plans to conduct proofs of concept for EXAONE Tabular in 3 areas in the second half of this year: manufacturing, bio and healthcare, and finance, and link them to actual business.

LG AI Research said it also achieved world-class forecasting accuracy on GIFT-Eval, a global leaderboard developed by Salesforce, to evaluate time-series forecasting performance across industries.

Time-series data refers to data whose values change over time, such as product demand, raw material prices, electricity use and stock prices. Time-series forecasting is a core area in industry, but differences in data characteristics and patterns of change across industries have been cited as a technical challenge, making it difficult for a single general-purpose AI to forecast the future for multiple industries.

EXAONE Forecast, LG AI Research’s Time Series Foundation Model, ranked first in the zero-shot category, which predicts unfamiliar-domain data accurately without additional training.

LG AI Research said it also ranked second in the agentic AI category, where AI independently analyses data and performs forecasting in real working environments, proving both its potential use in industrial sites and its general applicability across industries.

EXAONE Forecast was trained on real time-series data from industries including internet, sales, energy, economics, healthcare and transportation, as well as 2 trillion virtual time-series data points that simulate the real world.

LG AI Research said EXAONE Forecast can use a single model to predict time-series across different industries such as finance, energy, manufacturing and healthcare, giving it a broader range of applications and higher development efficiency than previous approaches that required repeatedly developing separate forecasting models for each industry.

LG AI Research is expanding the application scope of EXAONE, centred on domain-specific foundation models optimised for industrial site data and work characteristics.

LG AI Research released EXAONE Path, a Pathology Foundation Model, in 2024. It is now developing a cancer agentic AI that diagnoses cancer types and supports decision-making by medical specialists.

LG AI Research is also expanding the types of data EXAONE can handle. It is also developing a Robot Foundation Model that understands visual information and controls robots so they can perform tasks in the physical world.

Soon-young Lee (이순영), head of the Data Intelligence Lab at LG AI Research, said LG’s competitiveness lies in data accumulated across diverse industries and deep understanding of industrial sites. She said LG will continue to expand a group of foundation models specialised for multiple industries to create a new axis of competition in the global AI market.