LG EXAONE Tops Industrial AI Global Rankings

LG's independently developed artificial intelligence, 'EXAONE', has emerged victorious in global industrial AI rankings, outperforming tech giants such as Google and Alibaba. The AI consists of two key models: 'EXAONE Tabular', which excels in predicting tabular data, and 'EXAONE Forecast', which accurately forecasts future data. Both models achieved state-of-the-art (SOTA) performance on global leaderboards in their respective categories.

'EXAONE Tabular' led the 'TabArena' competition, which gauges tabular data prediction abilities, securing the top spot in the categorical data prediction sector. This breakthrough model operates differently from existing methods that convert data into text for analysis; instead, it directly comprehends the structure of rows and columns and the relationships between data, a feature known as a 'tabular foundation model' (TFM).

The ELO score for 'EXAONE Tabular' is 1760, surpassing Google's recent 'TabFM' release at 1749. Importantly, EXAONE demonstrates remarkable adaptability and resilience, even with limited data or missing information, which can significantly reduce costs and time in industrial applications.

Following rigorous technical verification in manufacturing, bio, and financial sectors throughout the latter half of the year, LG plans to roll out EXAONE commercially. In the 'GIFT-Eval' leaderboard, which assesses time-series data prediction, EXAONE Forecast also claimed first place, outshining Google and Alibaba.

This model analyzes shifting data, such as product demand and raw material prices, and achieved the highest performance in the 'zero-shot' sector, accurately predicting first-encountered fields without additional training. LG AI Research bolstered EXAONE's stability by merging and training 2 trillion virtual time-series data points simulating real-world industrial scenarios with actual data.

Currently, EXAONE is put to use for raw material price prediction at affiliates like LG Electronics and LG Energy Solution, with recent expansion to over 8,000 stock market analyses in the Korean and US markets.

LG's strategy focuses on cultivating 'industry-specific foundation models' instead of general-purpose models, addressing unique challenges in various fields. Notable examples include 'EXAONE Path', aiding cancer diagnosis, and a 'robot foundation model' capable of visual information processing to control robots. Co-head of LG AI Research, Lim Woo-hyung, highlighted that this achievement demonstrates the potential of industrial data to provide Korean AI with a competitive edge.